It’s been a wild couple years for a team of builders like us.

Everyone feels it. The squeeze. The shifting priorities. The attempt to claim healthy, wealthy, and wise habits at a time when maybe attention is diverted to focus on others.

Software and hardware have seen massive shifts in 2026. We’ve learned some transformational lessons at a time when learning is hard-earned with globally fluctuating variables.

Companion.

The roots of the word are Latin: com (with) and panis (bread). A companion is someone you share bread with. I’m from ’90s. Bread means money.

So when I think of the Companion Core Computer, I don’t picture an AI avatar or another chatbot. I picture a hardware appliance that works. Something that earns its place, pulls its weight, and helps you make bread.

Own your intelligence when you turn your computer into a personal companion (device) that remembers your data and understands you in context.

Companion Memory keeps what you import and capture, and your agents can recall it on demand.

Connect it to a local agent like OpenClaw or Hermes.

The agent also gets Hub’s tools, and can work with MCP servers from the Hub app store.

The Hub for Your AI Activity

Taking the headache out of hosting!
No seat limits. No surveillance. You control the keys.

Save time with easy

Special SKUs only available on Kickstarter now! Get a Core 64 with 1 or 2 TB of storage. My favorite, the Core 128 with 2 TB, or the new 192 with 2TB.

Or, bring your own hardware - the server-on-a-stick option is 50% the price we will list it for in the store here for a limited time and it offers everything you need to launch Companion Hub with local models already downloaded and a 2 year Pro subscription for free.

Free accounts of Hub offer everything users need to get going and reach three apps remotely. Pro increases the number of installed apps you can reach remotely without extra effort on your side.

Everyone who pledges will recieve an invite to a private Backers Discord channel. There, we will share prompt recipes, skills, tutorials, and host conversations.

We hope you’ll join us on this journey. It’s going to be a wonderful adventure.

A developer copies a raw client schema into a chat window. Four seconds later, a cleaned-up JSON structure comes back. No warning appears. Nothing slows them down.

That moment explains most privacy concerns about using ChatGPT at work. Chat interfaces are built to remove friction, and they remove the pause where people would normally ask whether data should leave the building. Because the prompt box looks like a messaging app, workers treat proprietary architecture like a routine support ticket, and the habits they apply to email attachments or shared drives never kick in.

The interface rewards speed and quietly penalizes hesitation. Employees know the security policy. So, does why the policy keeps losing to the interface?

Are we really so easily persuaded by product design?

Should employees worry about privacy when using ChatGPT at work?

Yes. It's very difficult to know where inputs go. Analyses of user discussions show that enterprise concerns center on what data enters the model, followed by fears that it will be shared with third parties or retained for future training [8]. For anyone handling client records or internal architecture, those are concrete fears.

Worry, however, rarely changes behavior. Attitudes toward exposure vary widely, but adoption follows a "chat first, worry later" pattern that casual interface design reinforces [16]. A finished draft on screen beats a governance document in a shared folder. Performance reviews measure output, not data hygiene, so the momentum of getting the task done wins over abstract compliance risk.

What data collection fears drive resistance to corporate AI tools?

The core fear is that public AI tools act as data sinks that sidestep existing governance, with no way to verify how inputs are handled. Employees paste internal documentation and customer personally identifiable information into these tools because the chat box treats unstructured input as a feature, not a risk [9].

Traditional controls weren't built for this. Data loss prevention tools are tuned to catch file uploads and outbound attachments, not text typed or pasted into a browser prompt. Engineering and finance teams can leak sensitive data without tripping standard network monitoring.

There's also a longer-tail risk. If submitted data ends up in training, research on membership inference and extraction attacks shows that models can sometimes reveal information about the data they learned from [9]. How much that applies depends on the plan and settings in use, and that uncertainty is the problem. Most employees can't tell which rules apply to the tool in front of them, so they can't judge the real cost of the convenience. (For a closer look at how this plays out in one industry, see AI in Finance Is Powerful, But It Can Expose What Matters Most.)

How should organizations assess employee perceptions of generative AI before deployment?

Start by measuring, not mandating. Written guidelines alone don't stop leakage that's driven by interface design [NEEDS SOURCE]. Nobody rereads an acceptable use policy while racing a deadline, and in the moment the cost of following protocol usually feels higher than the risk of skipping it.

A useful assessment does three things:

• Measures sentiment before rollout. Find out which teams see AI as a productivity multiplier and which see it as a surveillance threat.
• Pairs usage data with training. Telemetry shows what people actually do; behavioral training addresses why they do it [9].
• Tracks continuously. Ongoing monitoring and transparent communication reduce the anxiety linked to falling productivity and turnover risk [13].

Programs fail when they ignore how much people value speed. Teams need tools that are as fast as the ones they're being asked to give up.

Does the perceived risk outweigh the productivity benefits for workers?

Rarely. Asking AI to draft an email or refactor code feels like asking a helpful colleague, not like sending data to a cloud service under broad terms of service. Employees treat these tools as easy work buddies that need no technical expertise [5].

That's why shadow AI persists. Nearly two-thirds of organizations suspect widespread unsanctioned AI use, yet usage stays high because compliant alternatives cost more effort than they seem worth [8]. When the approved path means manual redaction or an IT approval queue, the unapproved one wins. The productivity gain is immediate and visible; the privacy risk stays invisible until an audit. Any governance plan that ignores this tradeoff will fail, however strict the written policy.

What shapes employee trust when sharing sensitive work with ChatGPT?

Trust erodes when privacy controls are hard to find or hard to understand [9]. Most staff can't parse a service agreement to confirm whether their inputs train future models, so they're left relying on vendor promises, and every high-profile leak makes those promises easier to doubt.

Three things rebuild trust [10]:

• Clarity about job security and how AI will be used.
• Predictable data handling that employees can actually understand.
• The option to run equivalent workloads on hardware they control.

People comply more when they can see where their inputs go and confirm that nothing leaves their environment. Local inference removes the uncertainty entirely, letting teams use AI fully on sensitive work. (If cost is the objection, see How Much Does a Home AI Server Cost vs ChatGPT? A 3-Year Comparison.) Trust comes from architecture you can inspect, not from marketing.

<! data-preserve-html-node="true"-- SECTION: example-mapping-a-shadow-ai-exposure-pathway -->

Example: Mapping a Shadow AI Exposure Pathway

Here's how a security team can find and close a leak in practice.

1. Detect. Watch for outbound traffic to known AI inference domains that lacks enterprise authentication, and flag prompts containing unredacted customer PII or architecture details [11].
2. Attribute. Compare query volumes by department against vendor retention and usage logs to see which workflows generate unsanctioned traffic [12]. Engineering often shows up through high-frequency coding prompts; support teams through pasted customer conversations.
3. Fix. Apply least-privilege controls at network egress, then give each team a local AI alternative with equivalent capability, so blocking the public tool doesn't just push usage further into the shadows.

The symptom reveals the mechanism, and the mechanism dictates the fix.

Closing the Gap: Governance That Works With People

Telling employees to stop using AI without offering a real alternative creates a compliance paradox that breeds resentment and hidden leakage. Continuous monitoring and transparent communication help [13], but the lasting fix is architectural: give people tools that are fast and private, so they never have to choose [14].

That's what local AI hardware makes possible. A Companion Core runs Hub apps and AI models on a machine the organization owns, with no per-seat subscription, and prompts and files stay on that hardware by default [15]. (For how local and cloud can work side by side, see Designing the Hybrid Future: How Local and Cloud Computing Work Together.)

Back to our developer. They paste the same client schema into a model running on their team's own Core. The cleaned JSON comes back just as fast, and the schema never leaves the office.

Governance catches up not by policing behavior, but by building systems that match how work actually gets done.

Believe in local AI? Back Companion Core on Kickstarter. Early bird discounts start October 13.

Back Companion Core on Kickstarter · Explore Companion Core

The answer arrives in seconds, in a calm and familiar voice. It feels effortless. But behind that smooth interface, each question travels across the internet, is processed on a machine in a distant data center, and comes back as a reply.

We rarely think about how it works, as long as it works.

Where does your conversation actually go?

Privacy at home usually gets framed as a policy choice, but it really comes down to plumbing.

A cloud AI service keeps your conversations on its servers, not in your home. That arrangement buys convenience, speed, and free tiers, but it also means the retention period, training defaults, and access rules are set by someone else.

When those settings change, or when a new regulation changes how long chats are kept, your household has to react rather than decide.

With most cloud assistants, what you type leaves your home immediately, travels to a provider's data center, and is processed and stored there. Companies typically use that input for three quiet purposes:

1. Keeping the service running and secure
2. Improving their models, often on by default unless you opt out
3. Occasionally sending samples to human reviewers for quality and safety checks

None of this means someone is actively listening. It is simply how cloud services work at scale.

What changes when you run AI on your own network?

Moving AI onto your own network changes the plumbing without changing how you use it.

A local server sits under the desk, connected to the same router as the television. When you run a model there, your question never leaves the house. The hardware reads the text, keeps the context, and writes the reply right there.

There is no trip to a data center and no outside company setting the rules for your conversation. You still get capable models that understand context, but the privacy boundary is physical rather than contractual.

Does everything need to run locally?

No. Not every task belongs behind your firewall.

Live weather, map directions, and software updates still need a connection to outside services. The practical split looks like this: keep personal notes, drafts, photo libraries, and family calendars on your local machine, and send only public, non-sensitive tasks to outside providers.

You choose which apps can reach the internet and which stay entirely within your home network. Setting up that private zone once means you don't have to re-check privacy settings every time a cloud service changes its terms.

Setup guides and architecture diagrams are in the Companion documentation.

What hardware do you need to run AI models locally?

It depends on the size of the models you want to run. The deciding factor is memory: VRAM on a graphics card, or unified memory on machines that share it between processor and graphics, such as Apple Silicon Macs and AMD Ryzen AI systems.

A private AI server for the whole family

Companion Core is built to be that box under the desk. It arrives ready to run, with an AMD Ryzen AI Max+ 395 and up to 128GB of unified memory, enough for 70B-class models on the device itself. Every Companion Hub app and local AI model runs at home, so homework questions, family photos, and personal notes stay on hardware you own.

Companion Core is on Kickstarter now, at the lowest price it will ever have:

• Core 64 — $3,200, Kickstarter only
• Core 128 — $5,400, Kickstarter only

Kickstarter pricing ends with the campaign. After that, the Core sells at a higher price on our website.

Back Companion Core on Kickstarter →

Want to try it first? Companion Hub is free and runs on the Mac, Windows, or Linux computer you already have.

Yes, for most people, running AI at home is very achievable in 2026. It does come with real hardware and budget considerations, and memory prices have been climbing, so it's worth knowing what you need before you buy anything. This guide explains what hardware matters, whether your current computer qualifies, and how to get started without writing a single line of code.

One note before we begin: running a model locally keeps your prompts on your own machine, but the app around it can still make network connections for updates or telemetry. Privacy depends on how you set things up.

What kind of hardware do you actually need?

The single most important number is memory. To run a model comfortably, your graphics card's memory (VRAM), or your computer's shared "unified" memory, needs to hold the model plus some headroom for the conversation itself.

Consumer GPUs vs. enterprise hardware for large models

For everyday chat, writing, and coding help, modern consumer graphics cards and Apple Silicon handle models up to about 30 billion parameters well.

Larger models are where memory runs out. A 70B model needs roughly 40GB of memory at 4-bit quantization and about 75GB at 8-bit, before leaving room for long conversations. The largest single consumer graphics card today has 32GB, so most home PCs can't fit a 70B model on one card.

You don't need data-center hardware to get there, though. Your options are:

• Two or more graphics cards working together, which takes more setup and more power.
• A high-memory Mac with 64GB or more of unified memory.
• A unified-memory machine built for it. Companion Core uses an AMD Ryzen AI Max+ 395 with up to 128GB of unified memory, enough to run 70B-class models on the device itself.

Is a local AI setup worth the cost?

It depends on what you spend now and what hardware you need. Be wary of anyone promising a fixed payback period.

If you pay around $20 a month for an AI subscription, that's $240 a year. Repurposing a computer you already own, or adding a used graphics card for a few hundred dollars, can pay off within one to two years. A dedicated machine costing thousands won't pay for itself on a single $20 subscription alone.

The math improves quickly as your spending grows. If you pay for several AI tools, for a premium tier like ChatGPT Pro at $200 a month, or for heavy pay-per-use API calls, a dedicated machine can pay for itself in one to two years. Heavy API users, such as people analyzing files daily or running long conversations, feel the savings fastest because local use has no per-token charges.

Beyond the money, local AI ends subscription fatigue and keeps your prompts and files on your own hardware, as long as you check your app's update and telemetry settings.

For a worked example with real numbers, see our post When Ownership Beats Subscriptions: The Family Tech Math.

The easiest way to get started without coding

You don't need to touch a terminal. These free desktop apps let you browse models, download them with a click, and start chatting:

For most beginners, LM Studio is the friendliest first step.

If you want more than chat, such as private file storage, photo backup, and a catalog of one-click local apps alongside your AI models, Companion Hub is free and runs on Mac, Windows, and Linux.

Want hardware built for this?

If your current computer falls short, or you want a quiet machine that runs 70B-class models without a multi-GPU build, Companion Core arrives ready to go. It runs every Companion Hub app and local AI models on an AMD Ryzen AI Max+ 395 with up to 128GB of unified memory.

Companion Core is on Kickstarter now, at the lowest price it will ever have:

• Core 64 — $3,200, Kickstarter only
• Core 128 — $5,400, Kickstarter only

Kickstarter pricing ends with the campaign. After that, the Core sells at a higher price on our website.

Back Companion Core on Kickstarter →

Frequently Asked Questions

What is the minimum GPU I need? A graphics card with at least 8GB of VRAM can run small 7B–8B models. On a Mac, 16GB of unified memory is a comfortable minimum; 8GB works only for the smallest models.

How much storage do I need? Plan for 50–200GB of free SSD space. A small model takes about 5GB, while a 70B model takes about 40GB, and most people keep several models installed.

Can my home PC run a 70B model? Usually not on a single graphics card. A 70B model needs about 40GB of memory at 4-bit quantization. You'll need multiple GPUs, a Mac with 64GB or more of unified memory, or a machine like Companion Core with up to 128GB of unified memory.

Does local AI guarantee complete privacy? No. Running locally keeps your prompts and files off a company's servers, but the app may still check for updates or send usage data. Review your app's settings to see what is switched on.

How much does Companion Core cost? On Kickstarter, Core 64 is $3,200 and Core 128 is $5,400. Those prices are Kickstarter-only; the Core will cost more on our website after the campaign. Back it on Kickstarter.

Ready to take control of your data?

Stop renting your AI and start owning it. Check your current computer against the tiers above, install a free app like LM Studio or Jan, and try a model today. The hardware barrier is lower than it has ever been, and private, subscription-free AI is within reach for far more people.

The monthly statement hits the inbox on a Tuesday morning. Cloud storage, AI chat seats, and a creative suite that raised its price last year all sit in a row. Canceling any of them means exporting files, finding them a new home, and hoping nothing breaks. So the question hangs over every recurring bill: is renting forever cheaper than owning the hardware?

How much do family tech subscriptions actually cost?

More than most households add up. A 2TB Google One plan is $9.99 a month, about $600 over five years, and that is the cheap line item. One ChatGPT Plus seat is $20 a month. Adobe Creative Cloud Pro is $69.99 a month on an annual plan, or $839.88 a year, after Adobe raised it from $59.99.

Subscriptions compound because they rarely pause, they charge per seat, and they bundle features you already pay for elsewhere. The first step is simply listing what you pay for. The math changes once a household has several users and several overlapping tools, because that is where cloud providers start charging premium tiers.

Is a family server cheaper than a SaaS stack over five years?

For households with heavy data and several paid AI or creative tools, usually yes. Cloud storage and seat fees never stop, and they grow as your photo library and family grow. Hardware is a one-time cost that keeps working until a part fails.

For households with under 1TB of documents and one or two subscriptions, probably not. A basic NAS box costs a few hundred dollars before drives, and a $10-a-month plan takes a long time to pay that back. Ownership wins when the hardware replaces many services at once, not just storage.

Where is the break-even point?

You break even when the subscriptions you cancel add up to more than the hardware price plus the electricity to run it. The formula is simple:

Break-even (years) = Hardware price ÷ (Annual subscriptions replaced − Annual power cost)

Power is the part most comparisons skip, so here it is in full. A device drawing 300 watts around the clock uses about 2,628 kWh a year. At the U.S. average of 18.2¢ per kWh forecast for 2026, that is about $478 a year. At California's 33.35¢, it is about $876.

That is a worst case: real machines idle well below their full-load draw. Use your own utility rate, and the wattage your device actually averages, for an accurate number.

What hidden costs come with owning?

Your time is the biggest one. A DIY server means command lines, network settings, and updates, and for a non-technical family those hours can wipe out the savings. Plan for drive replacements every several years too.

Remote access is the other cost. Cloud apps sync across every phone and laptop by default, and rebuilding that at home usually means setting up a VPN or port forwarding. A managed appliance is worth paying for when it removes that work rather than handing it to you.

Does self-hosting really reduce spending?

It does when one machine does several jobs at once. A single box that runs local AI models, syncs documents, backs up photos, and streams media can retire several subscriptions, not just one.

A hybrid setup is often the smartest middle ground. Keep a small cloud plan for the files you share every day, and move bulk photos, media, and backups onto hardware you own. Software prices have risen repeatedly in recent years; hardware locks in its cost the day you buy it.

Which devices give the best long-term return?

The return scales with how many services a device replaces, not how much storage it holds.

An entry-level NAS is the cheapest way in, but it lacks the memory and compute to run modern AI models, so you keep paying for cloud AI. AI workstations can run those models, but most are aimed at developers, not households.

Example: how fast does a Companion Core pay for itself?

At the Kickstarter price, a household replacing about $2,500 a year in subscriptions breaks even in roughly three years. The table assumes the worst case on power: a full 300 watts, all day, every day.

Companion Core 128 at $5,400. Real-world power averages below 300W, so actual break-even comes sooner.

To find your row, add up what your household pays for AI chat seats, cloud storage and photo backup, AI image tools, and creative software you could move to local apps. Two adults and a teen on ChatGPT Plus is $720 a year before anything else. Or use the calculator below to plug in your own subscriptions and electricity rate.

The verdict: when ownership becomes the cheaper stack

That Tuesday statement can become a one-time purchase. Once a family pays for four or more overlapping subscriptions, or wants AI that does not charge per seat or send data off-site, owning the hardware wins.

Companion Core puts compute, storage, and AI in one device. It runs every Companion Hub app and local AI models on the device itself, so your data never leaves your home. Inside is an AMD Ryzen AI Max+ 395 with 128GB of unified LPDDR5X memory.

First-party Hub apps (Spellbook, Earth, Photogram, and Just In Case) are free to install on macOS, Windows x64, and Linux. Need remote access without setting up a VPN? Hub Pro adds extra secure tunnels for one flat price, not a per-seat fee.

Back Companion Core on Kickstarter

Kickstarter is the lowest price Companion Core will ever have. Prices on our website will be higher.

• Core 64: $3,200, Kickstarter only
• Core 128: $5,400, Kickstarter only

Backers in the first 24 hours also join a private backers-only Discord, with free workshop invites and lifetime access to our learning materials.

Back Companion Core on Kickstarter →

Prices as of October 2026. Sources: Adobe Creative Cloud pricing, NVIDIA DGX Spark price increase, EIA residential electricity outlook, electricity rates by state.

If keeping your thinking, your processes, and your clients' information private matters to you, our answer is yes.

Cost is a fair concern, though. Memory prices keep climbing, and once our Kickstarter campaign ends, Companion Core will cost more on our website. Timing matters, so let's look at it honestly.

Is now the right time for a non-developer to take on a piece of hardware? If you use AI often enough that the cost math works, and you handle client information sensitive enough that keeping it in-house matters, then yes. The technical barrier has dropped a long way. With Companion Hub, running on a Companion Core or on a spare computer with at least 16GB of RAM, free alternatives to popular cloud tools install in minutes.

The question carries a hidden assumption: that developer skills are the price of entry. In 2024, that was close to true. Setting up local AI meant a terminal, a config file, and a free weekend. In 2026, that assumption no longer holds for most of what a business owner needs AI to do.

This post answers the two questions you are probably asking, in order. Can you use this without technical skills? And will it help your business? It also covers what most local-AI articles skip: where a personal AI server is not worth it, so you can make the call for yourself.

What can a non-developer do with a personal AI server?

Companion Core servers ship ready to run. Everything you need is already installed. You create an account, choose the apps and models you want, and get to work.

From there, you can ask questions about documents, draft client proposals and emails, summarize meeting recordings, review code before it goes to a developer, and search across your private company files. The model reads everything on your own hardware, so client contracts and meeting audio never leave your office.

Here is what non-technical business owners are doing with it today:

• Document analysis and Q&A. Upload a contract, a lease, or a proposal and ask questions about it. The AI reads the document on your hardware, and no copy goes to an AI company's servers.
• Draft generation. First drafts of proposals, client emails, and marketing copy. For most everyday writing, today's open models hold their own against a ChatGPT Plus subscription.
• Meeting summaries. Transcribe a recording locally and generate a summary. The audio never leaves your network.
• Code review, without being a coder. If you work with developers, you can use local AI to read scripts, automations, or website changes before they go out.
• An internal knowledge base. Ask questions across your own contracts, reports, SOPs, and past projects without sending any of it to a cloud service.

Now for the honest limit. Cloud AI is still better for tasks that need live web search, the most current information, or the deepest reasoning, such as complex legal analysis or multi-step strategic planning. Most business owners settle into a hybrid: local AI for sensitive work, cloud AI for general research. That isn't a compromise. It's a sensible way to work, and knowing which task belongs where is the practical skill this post is meant to give you.

Do you need to be technical to set one up?

No. There is now a path that never touches a terminal. LM Studio installs like any desktop app and gets you chatting with a local model in about five minutes. Jan.ai looks and feels the most like ChatGPT. Ollama, once command-line only, now has a desktop app as well. Companion Core goes furthest: pre-configured hardware with a simple interface, built for people who would rather not think about servers at all.

LM Studio is the one most non-technical users should try first. It downloads as a standard application for macOS, Windows, and Linux. Open it and you get a searchable model browser: filter by task, search by name, and see how much memory a model needs before you download it. According to a DEV Community comparison from May 2026, you can have a working chat in under five minutes from first launch. No command line, no config files, no server to manage.

Jan.ai is the closest local tool to ChatGPT in look and feel. Downloads are point-and-click, the chat interface is easy to pick up, and it works fully offline once a model is installed. It is also open source, so anyone can audit it.

Ollama is the name you will see recommended most often. It was long a command-line tool built for developers, but it now has a desktop app for macOS and Windows. Its depth is still aimed at developers building on top of local models, so for a first experience, LM Studio or Jan is the gentler start.

Companion Core, running Companion Hub, goes a step beyond all three. The hardware arrives pre-configured. The Hub gives you one-click app installs with no Docker, no Compose files, and no terminal, and it was designed from the start for non-technical operators. It also offers something the others don't: a full self-hosted setup in one place, with AI models alongside file storage, a password manager, and productivity apps.

One honest caveat, because overselling ease is how people get burned: even with LM Studio, getting to a working setup takes more time and attention than signing up for ChatGPT. Much faster than in 2024, yes. Instant, no.

Does it pay for itself?

It depends on what you pay today. If you pay $20 a month for ChatGPT Plus, a budget mini-PC can pay for itself, but a Companion Core will not on AI costs alone. If you pay $200 a month for ChatGPT Pro, a Core 64 at its Kickstarter price pays for itself in about 16 months. If a team of five shares ChatGPT Business, a Core 64 pays for itself in a little over two years, and it handles everything else the Hub offers on top.

Start with the cases where the math is clearest. ChatGPT Pro at $200 a month is $2,400 a year. At the Kickstarter price of $3,200, a Core 64 covers that in about 16 months, and a Core 128 at $5,400 in a little over two years. A team of five on ChatGPT Business (formerly ChatGPT Team) at $25 per user, billed monthly, spends $125 a month, or $1,500 a year. A Core 64 matches that in about 26 months.

Here is the simplified version. The full three-year comparison, including electricity and hardware tiers, is in our detailed cost breakdown.

Payback times are before electricity. The Core 64 and Core 128 are $3,200 and $5,400 on Kickstarter; prices will be higher on our website after the campaign.

Read the table against your own usage rather than ours. The point is that the answer depends on what you pay and how many people share it, not on a single headline number.

So, plainly: if you are a light ChatGPT Plus user, a Companion Core does not pay for itself on subscription savings alone, and we won't pretend it does. A budget mini-PC would. The financial case for the Core is strongest when you pay for Pro, when you have a team, or when the confidentiality of your client data has real value. That is the next section.

If your numbers land in the Pro or team rows, the Kickstarter is the least you will ever pay for a Core. See the Kickstarter tiers →

What do you have to manage if you are not a developer?

Less than a traditional server, more than a cloud subscription. Day to day, you update models and apps with a click from the Hub dashboard in your browser. Occasionally, you restart the hardware or confirm your backups are running. Rarely, something like a drive failure or setting up outside access may call for a setup guide or a quick support ticket. It is closer to owning a NAS than running a server.

Here is the breakdown, by how often each task comes up and how much it asks of you:

• Ongoing, low-effort, all in your browser. Model updates are one click when a new version lands, and app updates work the same way. Available storage shows on the dashboard. Adding a new app is a click from the catalog.
• Occasional, and rarely technical. Now and then you restart the Core, the same as any device. If your home or office network changes, you adjust for it. From time to time you confirm your backups are running, which the Hub surfaces for you.
• Rare, but worth knowing about. A few things may need help: connecting a self-hosted app to an outside service, recovering data after a hardware failure, adding storage, or setting up access from outside your office network. That last one uses Tailscale or a similar service and means following a setup guide, not deep technical work.

The bottom line: managing a Companion Core is closer to managing a NAS than a server. It is not zero maintenance, and it is not a set-and-forget cloud service. It is well within reach of a non-technical business owner, as long as you expect that an unusual situation may occasionally call for a support ticket or a short consult.

When does client confidentiality make it worth it on its own?

When you handle information you are obligated to protect, the privacy benefit alone can justify the hardware. Lawyers, accountants, healthcare-adjacent professionals, and anyone under NDA hold client data where a leak has real consequences. Running AI locally means the contract, the audio, and the client files are processed on your hardware and never sent anywhere. That reduced risk never shows up in a subscription comparison.

For some business owners, privacy isn't abstract; it's a requirement. It matters most for:

• Legal professionals, where routing privileged material through cloud AI raises questions about attorney-client privilege. We covered this in how lawyers accidentally waive attorney-client privilege, and it is the sharpest version of this argument.
• Financial professionals handling client financials and tax information.
• Healthcare-adjacent work with HIPAA considerations.
• Anyone under NDA: client contracts, product roadmaps, acquisition discussions.
• Agencies and consultancies holding client source code, campaign strategy, and proprietary data.

In these cases, "running AI locally" means something concrete. A local model reads your contract on your hardware, and no copy goes to an AI provider. A local transcription model handles your meeting audio, and the recording never leaves your network. A local document search answers questions about your client files because those files are indexed on your own machine.

This extends the ownership argument past the subscription line. A single data breach, or a single client who no longer trusts you, can cost far more than the hardware. For a professional handling sensitive information, that reduced risk belongs in the "is it worth it" calculation, even though it never appears in a ChatGPT price comparison.

One caveat, to keep this honest: if you don't handle sensitive client data, the privacy benefit is real but not urgent. It is a reason, not the reason, and it deserves to be weighed that way.

Where is it not worth it?

On cost alone, it is not worth it if you use AI lightly and pay $20 a month. Local models also trail the leading cloud models on the hardest reasoning, have no live web access by default, and take more setup than signing up for a service. If your work centers on deep research or complex multi-step reasoning, cloud AI may still be the better tool for those tasks.

These are the areas where local AI is still harder than cloud AI, and any honest answer should include them:

• Model selection. You choose which local model to run. Cloud AI makes that choice for you.
• The capability ceiling. The best open models in 2026 are competitive for most everyday tasks, but they do not match the newest frontier models from OpenAI, Anthropic, and Google on the hardest reasoning.
• Real-time information. Local models have no internet access by default. They answer from what they were trained on, not from live search.
• Context length. Some local models handle less text at once than cloud models, though that gap has narrowed a lot in 2026.
• Integrations. ChatGPT's apps, connectors, and integrations are a broader ecosystem than what is available locally.
• Setup time. Even with LM Studio, getting to a working setup takes longer than signing up for ChatGPT.

In plain terms, you trade some top-end capability and some convenience for full control of your data and a different cost structure. For a business owner whose AI use centers on drafting, document review, and internal Q&A, that trade is small. For someone doing deep research or complex multi-step reasoning, cloud AI may still be the better choice for those specific tasks. That is the whole picture, benefits and limits together, which is the only kind of answer worth trusting before you spend money.

Ready to make the call?

If you've read this far and the answer is yes, because you use AI enough and your clients' data matters enough, Companion Core is the version built for someone who isn't a developer. The hardware arrives pre-configured, apps install in one click, and you never need to open a terminal. Inside is an AMD Ryzen AI Max+ 395 with up to 128GB of unified memory.

Companion Core is on Kickstarter now, at the lowest price it will ever have:

• Core 64 — $3,200, Kickstarter only
• Core 128 — $5,400, Kickstarter only

Kickstarter pricing ends with the campaign. After that, the Core sells at a higher price on our website.

Back Companion Core on Kickstarter →

Not ready for hardware yet? Install Companion Hub for free on a computer you already own and see how it fits your work.

Frequently Asked Questions

Do I need to know how to code to use a personal AI server? No. LM Studio and Jan.ai install like normal desktop apps and give you a working local AI chat in minutes, with no command line. Ollama now has a desktop app too, though its deeper features are aimed at developers. Companion Core goes further, with pre-configured hardware and a simple interface that needs no Docker and no terminal.

Will a local model be as good as ChatGPT? For drafting, document analysis, and internal Q&A, the best open models in 2026 are competitive with ChatGPT Plus. For the hardest reasoning, deep research, and live web search, cloud AI still leads. Most business owners end up using both and learning which task belongs where.

Is a personal AI server worth it if I only use AI occasionally? On cost alone, probably not. If you use AI lightly and pay $20 a month for ChatGPT Plus, the subscription math does not favor a Companion Core. The case gets strong if you pay for ChatGPT Pro, run a team, or handle client data where privacy has real value.

How much does Companion Core cost? On Kickstarter, Core 64 is $3,200 and Core 128 is $5,400. Those prices are Kickstarter-only; the Core will cost more on our website after the campaign. Back it on Kickstarter.

What is the easiest way for a non-developer to start? Install LM Studio on the computer you already own. It downloads as a standard app, shows a searchable model browser with memory estimates, and gets you chatting in about five minutes. If you want the full self-hosted setup with no terminal at all, Companion Core with the Hub ships pre-configured.

Can I keep using ChatGPT alongside a local AI server? Yes, and most business owners do. A hybrid setup uses local AI for sensitive client work and cloud AI for general research and the hardest reasoning. Knowing which task belongs where is the practical skill; you don't have to pick one side forever.

Works Cited

• Running Local LLMs in 2026: Ollama, LM Studio, and Jan Compared (DEV Community)
• LM Studio vs Jan 2026 (Kunal Ganglani)
• Jan AI Complete Guide 2026 (AI Made Tools)
• Jan vs LM Studio vs Ollama 2026 (Local AI Master)
• Local AI Server for Business 2026 (Compute Market)
• Personal AI Servers 2026 (Boostely)
• Best Private Personal AI Assistants 2026 (Vellum)
• Self-Hosted AI Guide 2026 (GitHub Gist, yalexx)
• Ollama launches a desktop app for macOS and Windows (AlternativeTo)
• ChatGPT for Work pricing: Business & Enterprise (eesel AI)
• OpenAI Pricing (OpenAI)
• Companion Intelligence

Adding Up the Cents and Sense

The first time I added up my app bill, I did it on the back of a receipt while waiting for a coffee I knew I shouldn't buy.

Soon I was going to have to make a choice: coffee or convenience. Storage, notes, a password manager, the video-call tool I used twice a month, the signing service I needed once a quarter, a credit-check service for an apartment I never rented. None of them felt expensive on its own. Together they were a car payment's worth of small charges I had stopped noticing years ago.

If you have ever done that math, you know the feeling.

Some of you know the next part. I added up what I was paying to keep creative tools on hand: Adobe, Houdini, Substance, Notch, World Creator, and probably more I've forgotten.

It's a common problem

You are not alone in this. Subscriptions are designed to be easy to start and easy to forget, and most households carry more of them than they would guess.

With a new solution

Here is what changed. A few years ago, the answer to "can I just run this myself?" was "yes, if you give up a weekend to learn." In 2026 that answer is mostly just "yes." The setup effort that used to scare people off self-hosting has largely disappeared, thanks to app managers like Companion Hub that put large catalogs of local apps one click away.

Hub is free and runs on the Mac, Windows, or Linux machine you already have. If you want a dedicated box built for it, Companion Core is on Kickstarter now at campaign-only prices.

Now the real question is not whether you can replace a cloud subscription. It is which one you start with.

This post is a map, not a sales pitch. It walks through the cloud apps you can replace with a self-hosted server, which of them are already packaged on Companion Hub for one-click install, what you would realistically save, and an honest section on what self-hosting still does not replace well. Bring your own list of subscriptions and check them off as we go.

How Much of Your App Stack Are You Actually Renting?

The average household pays for eight to twelve cloud app subscriptions.

Most people pay for four or five without noticing, roughly forty to sixty dollars a month. Stacked together, common tools like storage, notes, a password manager, video calls, and document signing can pass one hundred dollars a month if you pay for all of them

Ready for a server built for this?

When you want a dedicated machine instead of your everyday computer, Companion Core runs every Companion Hub app, plus local AI models, on one quiet box at home. Inside is an AMD Ryzen AI Max+ 395 with up to 128GB of unified memory.

It is on Kickstarter now, at the lowest price it will ever have:

• Core 64 — $3,200, Kickstarter only
• Core 128 — $5,400, Kickstarter only

Kickstarter pricing ends with the campaign. After that, the Core sells at a higher price on our website. Backers in the first 24 hours also join a private backers-only Discord, with free workshop invites and lifetime access to our learning materials.

Back Companion Core on Kickstarter →

Not ready for hardware? Download Companion Hub for free and replace your first subscription on the computer you already own.

Here is the stack a lot of people are renting, with monthly costs as of mid-2026:

Pay for all nine and you are at about $108 a month, roughly $1,296 a year, and close to $3,888 over three years. Most people do not pay for the full set. The point is not the maximum. It is how little it takes to cross forty or fifty dollars a month without ever making a deliberate decision to spend it.

What You Can Replace, Category by Category

Nearly every common category has a self-hosted answer in 2026: Nextcloud or Seafile for files, Immich for photos, AppFlowy or AFFiNE for notes, VaultWarden for passwords, Jellyfin for media, Jitsi for video calls, Documenso for signing, and Home Assistant for smart home. Most are packaged on Companion Hub for one-click install. Here is the map, category by category.

• File storage and sync (Google Drive, Dropbox, OneDrive): Nextcloud is the most feature-complete self-hosted cloud (sync, document editing, calendar, contacts). Seafile syncs faster for large libraries. Both install via Coolify on Hub, with no data egress fees.
• Photo backup (Google Photos, iCloud Photos): Immich is the leading self-hosted Google Photos alternative, with mobile auto-upload on iOS and Android, face recognition, shared albums, and a timeline view. Production-ready, free, and on Hub. This is the one most people are surprised actually works.
• Notes, docs, and knowledge base (Notion, Confluence): AppFlowy is an open-source Notion alternative (docs, boards, databases, wikis). AFFiNE covers Notion and Miro in one tool, Excalidraw handles whiteboarding, and Joplin does encrypted markdown notes. AppFlowy, AFFiNE, and Excalidraw are on Hub.
• Password management (1Password, LastPass, cloud Bitwarden): VaultWarden is a Bitwarden-compatible server written in Rust. It runs on Raspberry Pi grade hardware, works with every Bitwarden client, and is the most popular self-hosted password manager in 2026. One-click on Hub.
• Media server, video (Plex Pass, parts of Netflix): Jellyfin is a free, open-source media server with no premium tier and no remote-streaming paywall, unlike Plex, which moved remote streaming behind Plex Pass in 2025. Plex is also on Hub for existing libraries. Both are Hub-available.
• Music streaming (Spotify, Tidal, for owned music): Navidrome streams your owned collection to Subsonic-compatible apps on iOS and Android, and Koel adds a polished interface. Both via Hub. Be clear-eyed here: a self-hosted music server streams the music you already own. It does not replicate Spotify's 100 million track catalog or its discovery. If you have a large local collection the replacement is complete. If you use Spotify mainly to discover, it is not.
• Team chat (Slack, Teams, private Discord): Element runs on Matrix, with end-to-end encrypted chat, rooms, voice and video, federated. Mattermost offers a Slack-identical interface and is used inside Companion Intelligence. Element is on Hub, Mattermost is in the Hub ecosystem.
• Video conferencing (Zoom Pro, Meet): Jitsi Meet is fully open-source, with no time limits and no account needed for participants. Confirmed on Hub.
• Project management and databases (Jira, ClickUp, Airtable, Asana): Taiga does full project management (Scrum, kanban, sprints, issues). NocoDB is a spreadsheet-database hybrid, an Airtable alternative. Both confirmed on Hub.
• AI image generation (Midjourney, Firefly, DALL-E): ComfyUI is a node-based Stable Diffusion interface that runs locally on your GPU, production-ready and confirmed on Hub. The images stay local.
• AI workflows (Zapier, n8n cloud, Make): Flowise is a visual AI workflow builder for agents, RAG pipelines, and automation chains, run locally. Confirmed on Hub.
• Document signing (DocuSign, HelloSign): Documenso is an open-source DocuSign alternative with full e-signature workflows, an audit trail, and PDF support. Confirmed on Hub.
• Smart home control (Google Home, Alexa cloud, SmartThings): Home Assistant runs locally, integrates with more than 3,000 devices, and needs no cloud. Confirmed on Hub. Your home data stays on your home network.

<! data-preserve-html-node="true"-- IMAGE BLOCK: Inline category-map visual. A tidy grid with rows for File Storage, Photos, Notes, Passwords, Media, Video Calls, Signing, Smart Home. Left cell: cloud-rented tool glyph. Right cell: self-hosted Hub replacement glyph with a one-click marker. Duo-tone teal and near-black, clean technical table style. -->

Which of These Are Already on Companion Hub?

As of June 2026, Companion Hub includes one-click installs for AppFlowy, AFFiNE, ComfyUI, Coolify, Documenso, Element, Excalidraw, Flowise, Home Assistant, Jellyfin, Jitsi Meet, NocoDB, PairDrop, Plex, Taiga, and VaultWarden, among others. That packaging is the difference: no Docker, no reverse proxy, no weekend of setup.

This is what separates the list above from a generic self-hosting roundup. Most articles hand you twenty open-source projects and a link to each one's install docs. The hard part is what comes next: spinning up containers, configuring a reverse proxy, sorting out certificates, and hoping the upgrade path does not break. Companion Hub collapses that into browsing a marketplace and clicking install.

The confirmed Hub app list as of June 2026:

• AppFlowy, AFFiNE, Excalidraw (notes, docs, whiteboards)
• VaultWarden (passwords)
• Immich (photos)
• Jellyfin, Plex (media)
• Jitsi Meet (video calls)
• Element (chat), with Mattermost in the Hub ecosystem
• Taiga, NocoDB (project management and databases)
• ComfyUI (AI image generation), Flowise (AI workflows)
• Documenso (document signing)
• Home Assistant (smart home)
• Coolify (the deployment layer that installs Nextcloud, Seafile, and more)
• PairDrop, Keila, and partner apps, with Hermes and OpenClaw available through the Hub ecosystem

One caveat worth stating plainly: this list is current as of June 2026. The Hub store keeps adding apps, so check the live marketplace when you go to install. If something here is not listed, or something new is, the live store is the source of truth.

What Would You Actually Save?

A hobbyist who replaces five common subscriptions (Google One, Notion Plus, 1Password, Zoom Pro, and DocuSign) saves about fifty-five dollars a month, roughly nineteen hundred dollars over three years. That is one example mix. Substitute your own apps to get your number, and it excludes media, AI, and project-management savings.

Here is that example, spelled out:

Over three years, that is about $1,979 on just these five subscriptions, and it does not count media streaming, AI generation, or project management. To be clear, this is one mix chosen to illustrate the math, not a universal claim about your bill. The honest way to use it is to take your own list of subscriptions, find the equivalents above, and run your own total. Your number is the one that matters.

What Self-Hosting Does Not Replace Well

A few cloud services have no clean self-hosted match yet. Team email is doable but complex to run. Self-hosted music servers stream your owned library, not Spotify's catalog or discovery. Real-time document collaboration with outside people, full Adobe Creative Cloud parity, and complex enterprise Jira workflows remain partial replacements rather than drop-in ones.

1. Team email. Self-hosted email (with Mailcow, for example) is doable but complex to run and keep deliverable. For critical communication, most people should keep a managed domain-email solution rather than self-host it.
2. Spotify's catalog. Navidrome and Koel replace your owned music library well. They do not replicate Spotify's 100 million plus track catalog or its discovery algorithms. If discovery is why you pay for Spotify, a self-hosted server is not a full replacement.
3. Real-time collaboration with outside people. Nextcloud and ONLYOFFICE support live document collaboration, but the external person has to connect to your server. It works, with a bit more friction than a shared Google Doc link.
4. Adobe Creative Cloud. Krita, Inkscape, and GIMP cover most use cases, and DaVinci Resolve handles professional video editing for free and locally. The gap narrows every year, but for some professional workflows it is not yet zero.
5. Jira at scale. Taiga handles most project management well. For complex enterprise Jira workflows, Plane (also self-hostable) is a closer match than Taiga. Pick the tool to the workflow, not the other way around.

None of this is a reason to stay fully in the cloud. It is a reason to start where the replacement is clean and keep the few hard cases on managed services until the gap closes. Knowing the limits is what makes the rest trustworthy.

Where to Start

Pick one app you already pay for and install its Companion Hub equivalent, then run it for a week before moving the next one. VaultWarden (passwords) and Immich (photos) are common first installs: high value, low risk, and quick to set up. You do not need to migrate everything at once.

That is the whole trick to making this stick. People who try to replace their entire stack in one weekend tend to burn out and roll back. People who replace one app, live with it for a week, and then move to the next one tend to keep going until the monthly bill is a fraction of what it was. Start with the one that annoys you most, or the one that costs you most. Either is a good first move.

If this sounds like your setup, Companion Hub is where the one-click versions of every app above already live. You can browse the marketplace, pick your first replacement, and have it running today. Download Companion Hub and start with one app.

Frequently Asked Questions

Do I need to know Docker to self-host these apps? No. The apps in this post are packaged on Companion Hub for one-click install, so you do not need Docker, a reverse proxy, or command-line setup to run them.

Can I really replace Google Photos? Yes. Immich is a production-ready self-hosted Google Photos alternative with mobile auto-upload on iOS and Android, face recognition, shared albums, and a timeline view. It is available on Companion Hub.

Does self-hosting replace Spotify? Only partly. Self-hosted music servers like Navidrome and Koel stream music you already own from your server. They do not replicate Spotify's 100 million track catalog or its discovery algorithms.

How much can a typical person save? As an example, replacing five common subscriptions (storage, notes, passwords, video calls, document signing) saves about fifty-five dollars a month, near nineteen hundred dollars over three years. Your number depends on the apps you pay for today.

Are these self-hosted apps actually production-ready in 2026? Yes for the named categories in this post. The honest-limitations section lists the cases (team email, music discovery, external real-time collaboration, full Adobe parity, complex Jira) that remain partial replacements.

Works Cited

• 10 Self-Hosted Apps You Need to Try — netbird.io
• Best Self Hosted Apps in 2026 — Pinggy
• 50+ Open-Source Alternatives to Cloud Services — DreamHost
• I ditched Plex, Google Photos, iCloud for self-hosted alternatives — XDA Developers
• Spotify Premium Price Increase Jan 2026 — gamsgo.com
• 1Password Price Hike March 2026 — 9to5google
• Notion Pricing 2026 — notion.com/pricing
• Companion Intelligence — ci.computer
• Companion Intelligence Hardware — ci.computer/hardware

A home server used to mean case modding and weekends spent troubleshooting power limits. Today it means pairing purpose-built hardware with software that does the heavy lifting. Here's how to move from cloud rentals to local compute.

Why Local AI Makes Sense

You own it instead of renting it

Cloud AI tools work until your budget doesn't. Stack a few premium subscriptions and you're paying every month for something you'll never own.

Companion Core is a one-time purchase. On Kickstarter, Core 64 is $3,200 and Core 128 is $5,400. These prices won't return once the campaign ends; the Core will cost more on our website. After that, your main ongoing cost is electricity.

If you rely on AI daily, that math turns in your favor fast. A cheaper $1,500–$3,500 server can run Companion Hub too. But for the best experience with Companion Memory, you'll want the memory bandwidth, storage, and RAM the Core is built around.

Privacy is an engineering choice, not a settings page

Local-first is a design constraint, not a marketing line. On the Core, models run on the device, and your prompts, files, and memory graph stay on your hardware by default. You aren't hunting for privacy toggles in someone else's dashboard; everything runs behind your own firewall. Remote access and public app URLs are opt-in, and you decide what leaves your network.

What You Need

The software: Companion Hub

You don't need to write orchestration scripts. Companion Hub is free, cross-platform software that turns a Mac, Windows, or Linux machine into a private AI server. It handles container management, local model routing, and agent coordination without the command line.

Hub includes an app store where developers publish one-click installs of local-first alternatives to popular cloud tools. Every app runs in its own container with dependency isolation and automatic updates. If an app breaks, it fails inside its container and the rest of your system keeps running.

The hardware: Companion Core

Hub runs on equipment you already own. Push past 13B-parameter models on typical hardware, though, and you'll hit memory ceilings or thermal throttling. That's where a purpose-built machine pays off.

Core is built on AMD's Ryzen AI Max+ 395 with a Radeon 8060S GPU, sharing up to 128GB of LPDDR5x unified memory. There's no PCIe bottleneck between CPU and GPU and no discrete-GPU driver to fight. It's tuned for sustained inference, stays quiet with a single Noctua fan, and the 128GB model runs 70B-class models on-device with llama.cpp. When you need more, 5GbE and USB4 let you link multiple units over RPC for larger or parallel workloads.

See Companion Core on Kickstarter →

Your First Setup

From box to running apps takes minutes:

1. Start up. Boot a Core, which ships with Companion OS preinstalled, or install Hub on your own machine.
2. Pair your hardware. Create a free Hub account at hub.ci.computer. It generates a six-character, single-use code that links your machine securely.
3. Run the setup wizard. Choose a default local model (Ollama comes ready to go), set up owner-only private access through Tailscale, and optionally turn on Cloudflare routing for HTTPS app addresses you can reach away from home.
4. Install apps. Deploy apps like Spellbook, Earth, Photogram, or Just In Case with one click. Each one pulls its own dependencies and shows up in your Hub dashboard.

No YAML unless you want it. The wizard handles networking, model downloads, and resource allocation so you can focus on what you're building.

Is a Home AI Server Right for You?

If you only do light prompt testing, cloud APIs may cover you. But if you chain agents, feed in personal context, run vision or multimodal models daily, experiment with agent harnesses, or keep hitting rate limits, local compute pays off quickly. You get predictable performance, full control of your data, and no vendor lock-in.

Take the First Step

You don't need to over-plan this. Hub is free and runs on what you already have, so start there and see how local-first AI works on your own machine.

Download Companion Hub — free

Back Companion Core on Kickstarter

When you're ready for hardware built for the job, the Kickstarter campaign is the cheapest the Core will ever be.

• Core 64 — $3,200: local AI, storage, and every Hub app on one quiet box
• Core 128 — $5,400: 128GB of unified memory for 70B-class models on-device

Kickstarter pricing ends with the campaign. After that, the Core sells at a higher price on our website.

Back Companion Core on Kickstarter →

When an AI model runs on your own machine, the prompts you send and the model's responses all stay on metal you can see (your computer hardware).

In a healthy system, nothing is sent to a company's servers. That is a strong privacy signal for your content, but it is not a guarantee the device is silent. The word "local" describes where the model runs, not the behavior of the whole app, so a local tool can still ping a server, check for updates, or reach the web for a search feature.

When someone says their AI runs locally, they usually mean it as a privacy claim, and most of the time they are right, just not for the reason or to the degree the word implies. The genuine win is real and it is structural: when the model lives on your machine, the thing you actually type never travels to anyone's server.

The slippage is that "local" gets used as a synonym for "private," full stop, and the people who say it rarely draw the line where the word stops being true.

If you already know about startup pings and telemetry, you do not need to be told that "local" can still provide feedback.

This post separates the real privacy win: your content staying home, from the small amount that can still leave, and it shows why that remainder is something you can check on your own machine rather than a policy you have to trust.

Does running an AI locally mean your data is private?

Running an AI locally makes your data more private, though not automatically, all of the time. The word "local" can be used to describe where the model runs, not the behavior of the whole app, so it is a strong signal rather than a guarantee. When it comes to a personal memory? That difference matters.

The genuine benefit of a generically labeled "local AI" product is worth stating, because it is the substance of the case. With the model running on your own computer, there is no API call carrying what you typed, and no conversation history sitting in someone else's account for later use. The thing most people are actually nervous about, a company reading, storing, or training on what they type, does not happen in that arrangement.

What remains uncertain is everything the surrounding app does around the model, and that is the part the rest of this post is about.

What does it mean when an AI "runs locally"?

An AI runs locally when inference, the model generating an answer, happens on your own device instead of a remote server. With local runtimes such as Ollama or LM Studio, your prompts stay on your machine, with no API call and no server-side log of what you typed or what came back. After a one-time download, the work happens on your hardware.

The distinction that matters is where your prompt lands. In the cloud, your prompt is the input to someone else's system, processed on their servers under their policies. Locally, the same prompt is the input to your own system, processed on hardware you control. That is the architectural difference under the word, and it is why "local" is a meaningful claim and not just a label.

Why is a local AI not automatically private?

A local AI is not automatically private because "local" describes where the model runs, not everything the surrounding application does. A tool can download the model weights locally and still phone home for telemetry, software updates, or optional cloud features. Local processing removes the main pipe your content would travel through; it does not promise the device is silent on the network.

The useful way to picture it: "local" is the room the model works in. The app is the building around that room, and a building can still have a door to the street. A tool that runs your model on-device and quietly sends usage data is not giving you the privacy you assumed you were getting. The good news is that there are only a few common doors, they are well understood, and most of them have a switch.

How can a local AI app still send data to the internet?

A local AI app can still send data in a few well-known ways. Most desktop GUI tools ping a server on startup, which reveals your IP and that the tool is running, even when your prompts never leave. Update checks and sometimes on-by-default telemetry are common. A built-in web search or retrieval feature sends the query you typed out to fetch results.

None of these send your full conversation to a training pipeline the way a cloud chatbot can, and most can be switched off. A startup ping says "this person is running this tool" and shows your IP; it does not carry what you asked. Likewise, an update check reaches a release endpoint to see if you are out of date. The one exception worth flagging is retrieval: if you turn on a web-search feature, the query you typed does leave your machine, because that is the only way to fetch results. That is a feature with a cost, and it is a cost you opt into.

What does a cloud AI like ChatGPT do with your data by default?

A cloud AI processes your prompts on its own servers, where the defaults matter. On ChatGPT's personal tiers (Free, Plus, Pro), the setting that lets OpenAI use your conversations to improve its models is on by default, and you opt out in Data Controls. Human reviewers can also access some conversations for safety and abuse monitoring under defined safeguards.

This is not an accusation against a company. It is the nature of hosted architecture: when the model lives on someone else's servers, your privacy depends on their settings, their defaults, and their incentives. You can turn the training setting off, but you have to know it exists and go find it.

In the cloud, the privacy you get is the privacy the provider chooses to give you by default.

What is the real privacy difference between local and cloud AI?

The real privacy difference is which question you have to answer. With a cloud AI, you ask "do I trust this company's policy, defaults, and incentives?" With a local AI, you ask "is this app on my machine reaching out, and can I stop it?" The second question is one you can answer yourself by checking settings and watching your own traffic.

That shift is the whole reason architecture beats policy. You cannot inspect a cloud provider's servers, read their internal logs, or confirm what they retain. You can inspect your own machine: open the app's settings, see what is switched on, and if you care to, watch what the device connects to. The cloud asks for trust because inspection is impossible. Local makes inspection possible.

Local vs cloud AI: what stays and what can leave?

Local and cloud AI differ most in what stays on your hardware. With a local model, your prompts and the model's answers are processed at home with no server-side log, and your conversations are not generated for a training set. What can still leave is connection-level activity such as a startup ping, an update check, or a web-search query, most of which you can switch off.

The table below lays out each line of the comparison, including whether you can verify it for yourself. aaaa

What are the privacy limits of running AI locally?

The privacy limits of running AI locally come down to scope. "Local" covers the model's location, not the whole app's behavior, so telemetry or a startup ping can remain. Web search and retrieval features re-open the network for the query you type. The operating system underneath makes its own connections. Verification is possible through settings and traffic, but it is not automatic.

These limits are the asset here, not a disclaimer, so they get full weight:

• "Local" describes the model's location, not the whole app's behavior. A locally run model can sit inside an app that still pings a server on startup, checks for updates, or sends optional telemetry.
• Features people want can re-open the network. A built-in web search or retrieval step sends your query out to fetch results, even when the model itself is local.
• Local does not silence the device underneath. The operating system and other software on the machine may still make their own outbound connections, independent of your AI app.
• "Private" still depends on checking. The architecture makes verification possible, settings you control and traffic you can watch, but most people will not actually inspect it. Local makes honesty possible; it does not make it automatic.

The accurate claim to keep is bounded. Say "your prompts and content stay on your hardware," not "nothing leaves your network." The first is the real, defensible win. The second is the kind of overclaim this post is warning against, and the audience that cares about this question will catch it.

Who should trust "it runs locally," and who should look closer?

Whoever wants the model itself to keep their prompts and content off a company's servers is a strong fit for "it runs locally," as long as they will check an app setting or two to confirm it. Anyone who reads "local" as "this device is now invisible on the network," and will not check anything, should look closer, because "local" alone overpromises for them.

The quick check is not hard. Open the app's privacy or network settings and look for update checks, telemetry, and any web-search or retrieval feature, which is where the switches usually live. If you want certainty rather than a reasonable default, watch the device's network traffic to see what it actually connects to. Either way, the work is yours to do, which is exactly the property the cloud cannot offer.

What still leaves your machine when you run AI locally?

What still leaves your machine is connection-level activity, not your conversation. Many desktop tools ping a server on startup, revealing your IP and that the tool is running. Update checks and sometimes on-by-default telemetry are common. A web-search or retrieval feature sends out the query you type if you use it. The operating system also makes its own outbound connections.

Named plainly, the remainder is short:

• A startup ping from many desktop GUI tools, which reveals your IP and that the tool is running, not your prompts.
• Update checks, and telemetry that is sometimes on by default in maintained software.
• The query from a built-in web search or retrieval feature, if you turn one on.
• Whatever the operating system and other software on the machine send on their own, independent of the AI app.

Each of these is a setting you can check or switch, or a layer below the app. None of them is your conversation going to a training pipeline. If you want the full network-level teardown of what a self-hosted box still emits, down to DNS, certificate checks, and the server name in a TLS handshake, that is the subject of a companion post on what data actually leaves your device when you self-host. Here, the awareness-level point is enough: the part that describes you stays home, and the rest is inspectable.

Why is privacy by architecture stronger than a privacy promise?

Privacy by architecture is stronger than a privacy promise because data cannot leave when there is no pipe for it to leave through. On a Companion Core, the AI runs on hardware you own, so your prompts, files, and the model's answers are processed at home with no path to a third-party AI service. That assurance comes from where the computer is, not from a policy you have to believe.

It follows from everything above. Local removes the pipe your content would have traveled through, and turns the small remainder into a startup ping, an update check, or an optional web lookup you can see and switch off. A "we take privacy seriously" promise asks you to trust a company whose defaults you cannot inspect. A structural arrangement asks you to trust the layout of your own machine, which you can. That is the difference Companion Intelligence is built on, and it is a difference in kind, not in degree.

If you are forming the mental model rather than buying hardware right now, that is the right place to be at this stage. The people comparing real local setups, and what actually leaks from them, are in the Discord, and that is the room where this stops being theory.

Running locally removes the part that describes you, your actual content, from the network, and turns the small remainder into something you can see and switch off. "Local" is the right direction and the right architecture. It is not a magic word that makes a device invisible, and anyone who tells you it is has borrowed the cloud's habit of overstating privacy. The honest version of "yes, local is more private" is the version worth keeping, because it is the one that survives the moment you actually check.

Frequently Asked Questions

Does running an AI locally mean my data is private? Mostly, but not automatically. When the model genuinely runs on your machine, your prompts and the model's answers are processed on your hardware and are not sent to a company's servers. "Local" describes where the model runs, not the behavior of the whole app, so it is a strong privacy signal for your content rather than a blanket guarantee.

What does "runs locally" actually keep on my machine? It keeps the sensitive core of the interaction on your device. With local runtimes such as Ollama or LM Studio, your prompts stay on your machine, with no API call and no server-side log of what you typed or what came back. After a one-time model download, the work of answering happens on your own hardware.

Can a "local" AI app still send data to the internet? Yes, in limited ways. Most desktop GUI tools ping a server on startup, which reveals your IP and that the tool is running. Update checks and sometimes on-by-default telemetry are common. A built-in web search or retrieval feature sends the query you type out to fetch results. None of these send your full conversation to a training pipeline, and most can be turned off.

Is local AI more private than ChatGPT? For your content, generally yes. On ChatGPT's personal tiers, the setting that lets OpenAI use your conversations to improve its models is on by default, with opt-out in Data Controls, and human reviewers can access some conversations for safety under defined safeguards. With a local model, your prompts are not sent to a company's servers in the first place, so the privacy depends on your machine, not a vendor's policy.

How can I check whether my local AI tool is sending anything out? Start with the app's own settings, where update checks, telemetry, and web-search or retrieval features are usually listed and switchable. If you want to be sure, watch the device's network traffic to see what it actually connects to. This is the structural advantage of local software: you can inspect your own machine, which you cannot do with a cloud provider's servers.

Yes, with honest caveats. For a non-developer entrepreneur in 2026 it is worth it when two things are true: you use AI often enough that the cost math closes, and you handle client data confidential enough that keeping it local matters. The technical barrier has dropped. Tools like LM Studio install in minutes, and Companion Core needs no terminal at all.

The question carries a hidden assumption: that developer skills are the price of entry. In 2024 that was close to true. Setting up local AI meant a terminal, a config file, and a free weekend. In 2026 that assumption is outdated for most of what an entrepreneur needs AI to do.

So this post answers the two questions you are asking, plainly and in order. Can you use this without technical skills? And will it help your business? It also does the thing most local-AI content skips: it tells you where a personal AI server is not worth it, so you can make the call for yourself.

What Can a Non-Developer Do With a Personal AI Server?

A non-developer can run document Q&A, draft client proposals and emails, summarize meeting recordings, review scripts before handing them to a developer, and search across private company files. The model reads everything locally, so client contracts and meeting audio never leave your hardware. Real-time web research still favors cloud AI.

Here is what non-technical entrepreneurs are doing with it today:

• Document analysis and Q&A. Upload a contract, a lease, or a proposal and ask questions about it. The AI reads the document on your hardware. No copy goes to OpenAI's servers.
• Draft generation. First drafts of proposals, client emails, and marketing copy. For most writing tasks, models like Llama 4 and Mistral perform comparably to ChatGPT Plus.
• Meeting summaries. Transcribe a recording locally and generate a summary. The audio never leaves your network.
• Code review without being a coder. If you work with developers, you can use local AI to read scripts, automation, or website changes before they go to your team.
• Client research and an internal knowledge base. Ask questions across your own contracts, reports, SOPs, and past projects through a local setup, without sending any of it to a cloud service.

Now the honest limit. Cloud AI is still better for tasks that need live web search, the most current information, or the deepest reasoning, such as complex legal analysis or multi-step strategic planning. What most entrepreneurs settle into is a hybrid: local AI for sensitive work, cloud AI for general research. That is not a compromise to apologize for. It is the intelligent position, and knowing which workload belongs where is the practical skill this post is trying to give you.

Do You Need to Be Technical to Set One Up?

No. The no-terminal path now exists. LM Studio installs as a normal desktop app and gives you a working chat session in under five minutes, with no command line. Jan.ai is the closest thing to a ChatGPT replacement. Companion Core goes furthest: pre-configured hardware with a consumer interface, no Docker, no terminal, built for non-technical operators.

LM Studio is the one to start with for most non-technical users. It downloads as a standard application for macOS, Windows, and Linux. Open it and you get a searchable model browser backed by Hugging Face: filter by task, search by name, see the estimated RAM a model needs before you download it, then click Download. Per the DEV Community comparison from May 2026, you can have a working chat session in under five minutes from first launch. No command line, no config files, no server to manage.

Jan.ai is the closest local tool to a ChatGPT replacement in look and feel. Point-and-click model downloads, an intuitive chat interface, and it is fully offline-capable once a model is downloaded. It is open-source, so the privacy-minded can audit it.

Companion Core, running the Companion Hub, goes a step beyond both. The hardware ships pre-configured and ready to run. The Hub gives you one-click app installation with no Docker, no Compose files, and no terminal, and it was designed explicitly for non-technical operators. It also adds something LM Studio and Jan do not: the full self-hosted stack, AI models plus file storage, a password manager, and productivity apps, in one interface.

The honest counterpoint matters here, because overselling ease is how people get burned. Ollama, which you will see recommended widely, is command-line first with no graphical interface. It is built for developers building applications on top of local models, and it is the wrong starting point for a non-developer. And even with LM Studio, getting from zero to a working setup still takes more time and attention than signing up for ChatGPT. Faster than 2024, yes. Instant, no.

Does It Pay for Itself?

It depends on what you pay now. If you pay $20 a month for ChatGPT Plus, a budget mini-PC pays for itself but a Companion Core does not on AI cost alone. If you pay $200 a month for Pro, a Companion Core pays for itself in under two years. If a team of five shares ChatGPT, Core financing of $128 a month is roughly break-even on the subscription alone, and adds everything else.

Start with the cases where the math closes cleanly. If you are paying for ChatGPT Pro at $200 a month, that is $2,400 a year leaving your account, and a Companion Core pays for itself in under two years while eliminating the ongoing fee. If you run a team of five on ChatGPT Team, that is $25 per user per month, $125 a month, $1,500 a year. Core financing through Affirm runs about $128 a month, which is roughly break-even on the AI subscription alone, and the Core then carries everything else the Hub offers on top.

Here is the simplified version. The full three-year cost comparison, with electricity and hardware tiers, lives in our detailed cost breakdown.

Read this against your own usage rather than ours. The point of the table is that the answer changes with what you pay and how many people share it, not with a single headline number.

So here is the honest math stated plainly. If you are a light ChatGPT Plus user, a Companion Core does not pay for itself on the subscription comparison alone, and you should not pretend it does. A budget mini-PC would. The financial case for the Core is strongest when you are paying for Pro, when you have a team, or when the confidentiality of your client data has a calculable value, which is the next section.

What Do You Have to Manage If You Are Not a Developer?

Less than running a server, more than a cloud subscription. Day to day it is click-to-update models and apps from the Hub dashboard, all browser-based. Occasionally you restart the hardware or confirm backups are running. Rarely, edge cases like recovering from a drive failure or setting up outside access may need a setup guide or brief support. It is closer to a NAS than a server.

The honest breakdown, by how often it comes up and how much it asks of you:

• Low-effort, ongoing, all browser-based in the Hub. Model updates are a click of Update when a new version lands. App updates are the same click. Available storage shows on the dashboard. Adding a new app is click-to-install from the catalog.
• Occasional, and rarely technical. Sometimes you power-cycle the Companion Core, the same as restarting any device. If your home or office network changes, you adjust for it. Now and then you confirm your backups are running, which the Hub should surface.
• Rare, but worth being honest about. A few things may need help: complex integrations between apps (say, wiring a self-hosted project tool to an outside service through an API), recovering data after a hardware failure, upgrading hardware such as adding storage, or setting up access from outside your local network. That last one uses Tailscale or an equivalent and means following a setup guide, not deep technical work.

The honest conclusion: managing a Companion Core is closer to managing a NAS device than managing a server. It is not zero maintenance, and it is not a set-and-forget cloud service. It is well within reach of a non-technical business operator, as long as you go in knowing that an edge case may occasionally call for a support ticket or a short technical consult.

When Does Client Confidentiality Make It Worth It on Its Own?

When you handle data you are obligated to protect, the privacy benefit can justify the hardware by itself. Lawyers, accountants, healthcare-adjacent professionals, and anyone under NDA carry client data where a cloud leak has real cost. Running AI locally means the contract, the audio, and the client files are processed on your hardware and never transmitted. That risk reduction does not appear in the subscription comparison.

For some entrepreneurs the privacy argument is not abstract, it is structural. It matters most for:

• Legal professionals, where routing privileged material through cloud AI raises attorney-client privilege questions. We covered exactly this in how lawyers accidentally waive attorney-client privilege, and it is the sharpest version of this argument.
• Financial professionals handling client financials and tax information.
• Healthcare-adjacent work with HIPAA considerations.
• Anyone under NDA: client contracts, product roadmaps, acquisition discussions.
• Agencies and consultancies holding client source code, campaign strategy, and proprietary data.

What "running AI locally" means for these cases is concrete. A local model processes your contract document on your hardware, and no copy is transmitted to OpenAI. A local transcription model handles your meeting audio, and the recording never leaves your network. A local document Q&A system answers questions about your client files because those files are indexed locally.

This is the economics-of-ownership argument extended past the subscription line. The cost of a single data breach or a broken client trust can exceed the price of the hardware by orders of magnitude. For a professional handling sensitive client data, that risk reduction is part of the "worth it" calculation, and it is a value that never shows up in a ChatGPT price comparison.

The caveat, to keep this honest: if you do not handle sensitive client data, the privacy benefit is real but it is not urgent. It is a reason, not the reason, and you should weigh it as such.

Where Is It Not Worth It?

It is not worth it on cost alone if you use AI lightly and only pay $20 a month. Local models also trail the frontier on the hardest reasoning, have no live web access by default, and take more setup than a signup. If your work centers on deep research synthesis or complex multi-step reasoning, cloud AI may still be the better tool for those specific tasks.

These are the things that remain harder than cloud AI, and they belong in any honest answer:

• Model selection. You have to choose which local model to run. Cloud AI hides that choice from you.
• The capability ceiling. The best local models in 2026, Llama 4 and Mistral, are competitive for most tasks but do not match the frontier capability of GPT-4o or Claude Sonnet on the hardest reasoning.
• Real-time information. Local models do not have internet access by default. They answer from training data, not live search.
• Context length. Some local models have shorter context windows than cloud models, though that gap has narrowed a lot in 2026.
• Plugins and integrations. ChatGPT's plugin and integration ecosystem is broader than what is available locally.
• Setup time. Even with LM Studio, getting to a working setup takes more time than signing up for ChatGPT.

The trade-off in plain language: you are trading some capability ceiling and some convenience for complete data sovereignty and a different cost structure. For a non-developer entrepreneur whose AI use centers on drafting, document analysis, and internal Q&A, that trade-off is minimal. For someone doing deep research synthesis or complex multi-step reasoning, cloud AI may still be the better tool for those specific tasks. That is the whole answer, upside and limit, which is the only kind of answer worth trusting when you are about to spend money.

If after all of that the answer for you is yes, that you use AI enough and your data matters enough, then a Companion Core is the version built for someone who is not a developer: the hardware pre-configured, the apps one click, the terminal nowhere in sight. See the Core and what it runs and decide.

Frequently Asked Questions

Do I need to know how to code to use a personal AI server? No. LM Studio and Jan.ai install like normal desktop apps and give you a working local AI chat in minutes, with no command line. Companion Core goes further, with pre-configured hardware and a consumer interface that needs no Docker and no terminal. Ollama still requires the command line, so it is not the right starting point for a non-developer.

Will a local model be as good as ChatGPT? For drafting, document analysis, and internal Q&A, the best local models in 2026 such as Llama 4 and Mistral are competitive with ChatGPT Plus. For the hardest reasoning, deep research synthesis, and live web search, cloud AI still leads. Most entrepreneurs end up using both and learning which workload belongs where.

Is a personal AI server worth it if I only use AI occasionally? On cost alone, probably not. If you use AI lightly and pay $20 a month for ChatGPT Plus, the subscription math does not close in favor of a Companion Core. The case gets strong if you pay for Pro, run a team, or handle client data where the privacy benefit has a calculable value.

What is the easiest way for a non-developer to start? Install LM Studio on the computer you already own. It downloads as a standard app, shows a searchable model browser with RAM estimates, and gives you a working chat session in under five minutes. If you want the full self-hosted stack with no terminal at all, Companion Core with the Hub ships pre-configured.

Can I keep using ChatGPT alongside a local AI server? Yes, and most entrepreneurs do. A hybrid setup uses local AI for sensitive client work and cloud AI for general research and the hardest reasoning. Knowing which workload belongs where is the practical skill, not picking one side forever.

Works Cited

• Running Local LLMs in 2026: Ollama, LM Studio, and Jan Compared (DEV Community)
• LM Studio vs Jan 2026 (Kunal Ganglani)
• Jan AI Complete Guide 2026 (AI Made Tools)
• Jan vs LM Studio vs Ollama 2026 (Local AI Master)
• Local AI Server for Business 2026 (Compute Market)
• Personal AI Servers 2026 (Boostely)
• Best Private Personal AI Assistants 2026 (Vellum)
• Self-Hosted AI Guide 2026 (GitHub Gist, yalexx)
• OpenAI Pricing (OpenAI)
• Companion Intelligence (ci.computer)

Microsoft retired its standalone OneDrive for Business plans in 2026 — if you were buying Plan 1 or Plan 2 as a standalone subscription, not bundled with Microsoft 365, your renewal window closes in January 2027 and the service shuts down in December 2029. Your files remain accessible until then, but you cannot renew a plan that expires before that date.

This post is specifically about standalone OneDrive for Business. Personal OneDrive accounts and OneDrive bundled with any Microsoft 365 subscription are not affected. If you're unsure which category you're in, check your Microsoft account before reading further.

For those who are affected: your migration window is open. Two self-hosted options — Syncthing and Nextcloud — let you move your files to hardware you own in under an hour.

Which OneDrive plans did Microsoft retire — and are personal accounts affected?

Microsoft retired its standalone OneDrive for Business Plan 1 and Plan 2 in 2026 — these are business plans purchased as standalone licenses, not part of a Microsoft 365 subscription. Personal OneDrive accounts and OneDrive bundled with any Microsoft 365 plan are not affected. If you're unsure which plan you have, check your Microsoft 365 admin center.

The retirement affects users who bought storage capacity directly from Microsoft without a broader productivity suite. These plans existed specifically for businesses and freelancers who needed cloud storage for work files — storage only, not the Teams-and-Exchange bundle that Microsoft 365 includes.

Microsoft stopped accepting new purchases for these standalone plans in May 2026. Existing customers whose plans renew before January 2027 can continue using the service until that billing cycle ends. After that, no renewal is possible regardless of billing type.

What happens to your files when OneDrive for Business standalone is discontinued?

Your files remain fully accessible until December 2029, when the service permanently terminates. The more immediate deadline is January 2027 — when existing plan renewals stop. If your plan is due to renew in 2027, you cannot renew it; you must migrate your files to another service before that date.

Microsoft has not published a specific data-deletion timeline beyond service termination, but data in a shut-down service does not remain accessible once the service goes dark. The practical window is tighter than December 2029 for most users: if your annual plan renews in 2027, your access ends at that renewal date. Monthly plans stop at the last billable cycle before January 2027.

The first action is the same regardless of which path you take: download your files to a local folder before the service ends. It is both a migration step and the most durable backup you can take right now.

Is Microsoft 365 Business Basic a good replacement for standalone OneDrive?

For a solo operator who only needed file sync, the replacement bundle costs more and adds apps you didn't ask for. If you already pay for Microsoft 365 for other reasons, this is a reasonable continuation. If you bought standalone OneDrive specifically because you didn't want the rest of the bundle, you're being asked to pay more for software you didn't need.

The honest version of this calculation depends on whether you have existing hardware. At 1TB for one user, $5 per month is not an expensive service. But it is a recurring payment, and if you already have a machine sitting underused, the alternative gets more interesting. The Mirage of the Cloud

What are the best self-hosted alternatives to OneDrive for Business?

The two strongest self-hosted replacements for standalone OneDrive are Syncthing and Nextcloud. Syncthing requires no server — it syncs files directly between your devices, is free, and takes minutes to set up. Nextcloud provides full web access, mobile apps, and file sharing, but requires a host machine and 30–45 minutes of setup.

These are not the same type of tool. Choosing between them depends on what you were actually using OneDrive for.

Path 1: Syncthing — peer-to-peer sync, no server needed

Syncthing syncs folders between devices using a direct encrypted connection. There is no central server, no cloud account, and no subscription. Each device holds its own copy of the synced files. It is open source, free, and available for Windows, macOS, Linux, and Android.

If you used OneDrive primarily to keep files synchronized between two or more machines you own, Syncthing replaces that directly — without a cloud intermediary.

Path 2: Nextcloud AIO — full replacement, server required

Nextcloud is a self-hosted file and collaboration platform. It includes a web interface, desktop and mobile clients, file sharing with external links, and optional apps for notes and calendar. You access your files via browser from anywhere, which Syncthing doesn't provide.

Nextcloud requires a host machine. The recommended deployment is Nextcloud All-in-One (AIO) via Docker, which handles the underlying container configuration. If Docker isn't installed yet, add 15–20 minutes for that step.

How long does switching from OneDrive to self-hosting actually take?

Switching to Syncthing takes about 15 minutes: install it on two devices, share the folder, and files sync immediately. Switching to Nextcloud takes 30–45 minutes after Docker is installed — longer because it sets up a full server environment, not just a sync client. Both paths start with the same first step: downloading your OneDrive files to a local folder.

Steps for both paths:

1. Download your OneDrive files to a local folder on your primary machine (time: varies by file volume and connection speed)
2. Choose your path: Syncthing for device-to-device sync without a server; Nextcloud for web access and a server you own
3. Syncthing: Install on both machines, add the shared folder, add the other device by ID, files begin syncing (~15 minutes from install to first sync)
4. Nextcloud: Install Docker if not installed, run the Nextcloud AIO setup command, follow the web-based wizard, point it at your data folder, done (30–45 minutes from Docker install)
5. Move the downloaded files into the synced folder

The initial file download is the longest step for most users and can run in the background while you complete the client setup.

What are the downsides of self-hosting your own file storage?

Self-hosting your file storage means you are responsible for uptime, backups, and hardware maintenance — tasks OneDrive handled automatically. Syncthing requires the source device to be online to sync; Nextcloud requires a server to stay running. The economic case for self-hosting is strongest when you already have a home server or machine to spare.

Neither Syncthing nor Nextcloud backs itself up. You need a separate backup strategy for the data on that hardware — an external drive, a rotation of local snapshots, or a cold-storage service. This is also true of OneDrive (Microsoft doesn't guarantee against accidental deletion), but you're managing it yourself now rather than having it handled invisibly.

What cloud storage still does better: redundancy across geographically distributed data centers, access from completely unfamiliar devices without setup, and near-zero recovery time if your local hardware fails. If you travel frequently and need your files from machines you don't control, self-hosting requires configuration each time. If you work from one or two machines in predictable locations, that friction is mostly absent.

Is there an easier way to set up self-hosted file sync without Docker?

Companion Hub installs containerized, one-click versions of self-hosted software on your own hardware — no command-line configuration required. If the Docker setup in the Nextcloud path looks like more friction than you want to take on, Hub's approach is to make that step disappear. You still own the hardware; the complexity doesn't follow you there.

You shouldn't need to know what Docker is to run your own file sync. Hub's interface handles that layer so the setup reads like installing any other application — you're choosing what to run, not configuring how it's packaged. The underlying software is the same open-source tools this post describes. The installation path is different.

You still own the machine, the data, and the sync process. The Return of the Local

The pattern here is recognizable: a standalone service gets retired, the bundle replacement costs more, and solo operators who paid for simplicity get pushed toward tools they didn't want. That's not new. What's changed is that the alternative — moving your files to hardware you own — has become genuinely practical in a way it wasn't a few years ago. The setup time is real but bounded. The ongoing cost is close to nothing if you have the hardware. Your migration window is open, and closing it takes an afternoon.

Stewardship Systems

Frequently Asked Questions

Works Cited

1. Microsoft Corporation. "OneDrive for Business Plan 1 and Plan 2 — Retirement Notice." Microsoft 365 Documentation. 2026.
2. Microsoft Corporation. "Microsoft 365 Business Basic." microsoft.com/en-us/microsoft-365/business/microsoft-365-business-basic. Accessed 2026-06-09.
3. Syncthing Foundation. "Syncthing Documentation — Getting Started." docs.syncthing.net. Accessed 2026-06-09.
4. Nextcloud. "Nextcloud All-in-One." hub.docker.com/r/nextcloud/all-in-one. Accessed 2026-06-09.
5. Nextcloud. "What is Nextcloud?" nextcloud.com/about. Accessed 2026-06-09.

What data leaves your device when you use apps on a self-hosted server?

Less than the cloud, but not nothing. Self-hosting keeps the content (files, photos, messages, AI prompts) on hardware you control. What still leaves is connection-level traffic almost any device makes: DNS lookups, certificate checks, time sync, and the server name in TLS handshakes, plus app update checks and any push or federation you turn on. Most of it is an architecture choice you can shrink.

The first time you point a DNS log or a packet capture at your own self-hosted box, you get a small, useful shock. You moved your files off the cloud, stood up your own server, told yourself the data stays home now. Then you watch the traffic, and the box is still talking. Not loudly, and not your photos, but it is reaching out: a lookup here, a certificate check there, a steady tick to a time server somewhere.

The instinct is to feel cheated, like the privacy promise was a lie. It was not. What you are seeing is the difference between two very different kinds of leaving, and once you can tell them apart, the picture gets clearer instead of scarier. This post is the honest wiring diagram: what leaves a self-hosted server, what stays, and why nearly everything that still leaves is a decision you can change rather than a policy you have to trust. Self-hosting did not make you invisible. It was never going to. The point is what it did do instead.

Does Self-Hosting Mean Nothing Leaves Your Network?

No, and any honest answer starts there. Self-hosting moves your content (files, photos, messages, AI prompts) onto hardware you control, so the high-value payload stops leaving. But the device underneath still does DNS lookups, certificate checks, and time synchronization, and the operating system may phone home on its own. Self-hosting silences the application data layer, not the network itself.

Before any app is involved, a machine on the internet emits a predictable set of outbound connections. These are wiring facts, not threats, and naming them is the difference between an honest map and a marketing claim.

• DNS lookups. Every time your server resolves a hostname (to reach an update server, a federation peer, a time pool, anything) it emits a DNS query. Unless encrypted, that query is visible to whoever runs the resolver and to anyone on the path. This is exactly why DNS logs are the standard tool for seeing what a device is doing: they reliably reveal telemetry endpoints, ad traffic, and chatty software.
• Certificate revocation checks. When a client opens an HTTPS connection, it may check whether the server's certificate has been revoked. The Online Certificate Status Protocol (OCSP) does this by querying a responder, typically over port 80, for certificates in the chain. That check tells the responder, often the certificate authority, which site you are connecting to and when.
• Time synchronization. Devices keep their clocks correct by reaching out to NTP servers. It is low-sensitivity traffic, but it is continuous, and TLS itself depends on it: OCSP responses are time-stamped, and a skewed clock makes valid certificates look invalid.
• Operating-system telemetry. The host OS and its components phone home on their own schedule, independent of your apps.

Pi-hole deployments make the scale of this background noise visible, and the honest detail is that the junk is concentrated, not evenly spread. Across a whole network the blocked share can look modest: one documented sample blocked about 8,081 of 268,793 queries in 48 hours, roughly 3 percent. But the average hides where it lives. A streaming or smart-TV device of the Roku class has been measured at roughly two-thirds of its DNS lookups being ads and telemetry, while a working desktop sits closer to 1 percent. Self-hosting an app does not silence the device or the OS beneath it. It silences the application-layer data flow, the part that contains your content.

What Do Self-Hosted Apps Send on Their Own?

Mostly two things: update checks and optional telemetry. A self-hosted app still wants to know if it is out of date, so it checks a release endpoint. Mature projects increasingly ship telemetry off by default and let you choose what, if anything, reports. The difference from the cloud is that here the outbound call is a setting you own and can watch, not the product itself.

The cleanest illustration is Immich, the self-hosted photo platform. Its automatic version check originally sent HTTP requests straight to github.com to read the releases of the Immich repository. The maintainers moved that check to their own dedicated endpoint, version.immich.cloud, specifically to avoid every instance contacting GitHub and to stop hammering GitHub with requests. The change shipped through public pull requests, numbers 18572 and 27450. That is the whole topic in miniature: the outbound call did not disappear, but a project that treats egress as a design concern got to choose where it goes and minimize what it reveals.

Telemetry follows the same pattern when a project is disciplined about it. Nextcloud is the reference: its telemetry is disabled by default, and the administrator chooses what an installation reports, if anything. (Nextcloud separately prompts you to turn on a feedback or usage-statistics app, the source of the recurring "Help Nextcloud" notifications people mention. That prompt is opt-in, not silent collection.) The contrast with cloud SaaS is the point. In the cloud, telemetry is the product and you cannot turn it off. In a well-behaved self-hosted app, it is a setting you own.

That word "well-behaved" is doing real work. Not every self-hosted project is disciplined about its outbound calls, and the only way to be certain is to watch the traffic with a DNS log or an egress monitor. Self-hosting at least makes that possible. You cannot packet-capture a cloud provider's servers. You can capture your own.

What Metadata Leaks Even When the Payload Is Encrypted?

The fact and target of a connection, even when its contents are sealed. Encryption protects what you send, not that you sent it or to whom. Unencrypted DNS reveals the hostname; the TLS SNI field has historically named the server in plaintext; the destination IP is visible to your ISP; and traffic timing and volume can profile activity. Encrypted DNS and Encrypted Client Hello shrink this, but coverage is still partial.

This is the part that survives "but it is all HTTPS," so it is worth walking each piece alongside the lever that shrinks it. Plain DNS sends the hostname in the clear, and encrypting it with DoH or DoT is necessary but not sufficient on its own, because other parts of the connection still reveal the destination. The lever is real but partial, which is the shape of this whole section.

The TLS SNI field is the sharper example. Even with TLS 1.3, which encrypts most of the handshake, the Server Name Indication that names the server you are reaching has historically been sent in plaintext in the ClientHello. Any network observer (your ISP, a network admin, an on-path actor) can read it and learn which domain you are contacting, even though they cannot read the content. This is not theoretical: South Korea began inspecting SNI to enforce site blocking in February 2019.

The fix is Encrypted Client Hello, which encrypts the SNI and closes that last handshake gap, and its status is the part you have to state with care. ECH is real and deploying. Cloudflare enabled it by default for its customers in late 2023, and Firefox and Chrome added support in October 2023. But coverage is still thin. As measured in 2025, roughly 4.2 percent of the top 100,000 sites and 9.2 percent of the top 1,000,000 advertised ECH support, and server support remains concentrated at Cloudflare. ECH has also been actively blocked in Russia, from November 2024, as well as in Iran and China. So the SNI fix exists, but assuming it is universally on would be wrong.

Two residues do not fully close. The destination IP is inherently visible to your ISP, and timing and volume patterns can profile activity even when names and contents are hidden. These are the irreducible cost of using a network at all. Why it still matters for the self-hosting case: the connection metadata that remains is largely metadata about your own infrastructure, your server reaching an update endpoint, a time pool, a federation peer, not a continuous stream describing your behavior to a company whose business is your behavior. The volume drops and the sensitivity drops. But "drop" is the accurate verb, not "eliminate."

What Leaves When You Add Push, Federation, and Lookups?

This is where a self-hosted setup quietly touches third parties. Mobile push generally routes through Apple (APNs) and Google (FCM) via a push proxy. Federation, by design, exposes the communication metadata between servers. And media apps fetch external metadata keyed to the title you add. Each is a feature with a cost, and each has an architecture that reduces what leaves.

A self-hosted app in isolation is easy to keep local. The egress shows up when you add the conveniences that made the cloud version attractive. Three are worth naming precisely.

Mobile push notifications are the single most common place a self-hosted setup touches Apple and Google. A self-hosted server cannot push directly to a stock mobile OS, so it hands the notification to a push proxy that forwards it to APNs or FCM, which deliver it to the device. Mattermost and Zulip both document running your own push-proxy for exactly this reason, and Nextcloud routes its mobile notifications through a Push Proxy service rather than letting every server contact Google and Apple directly. The nuance is real: by default the notification can pass through APNs or FCM, and the delivery token must be shared with them. The first lever is an ID-only mode, where only an opaque message ID transits APNs or FCM and the device fetches the actual content directly from your server, so Apple and Google see that a message exists but not what it says. The fuller answer on Android is UnifiedPush, with a distributor such as ntfy or NextPush: a decentralized, open protocol that lets you receive push without Google's FCM at all, pointed at a server you choose. The caveat survives here too. Some apps still reach fcm.googleapis.com even under UnifiedPush depending on the build, and on iOS, fully bypassing APNs is not realistic for most apps. So these levers reduce what Apple and Google see; they do not zero it out on every app.

Federation is a feature, and features have a cost. A self-hosted Matrix homeserver keeps your account and message content under your control, and end-to-end encryption protects the message bodies. But federation is, by design, servers talking to other servers, and the metadata of that conversation is exposed between homeservers: who talks to whom, when, from which domain, plus device IDs and timestamps. End-to-end encryption conceals the contents, not the social graph. If you only ever talk to people on your own server, none of this leaves. The moment you federate, the metadata does. That is a deliberate trade worth naming plainly, rather than implying that end-to-end encryption makes a federated server fully private.

External metadata lookups are the quietest of the three. Self-hosted media apps often enrich your library from public databases. Jellyfin, for example, fetches movie and show metadata and artwork from providers such as TMDB, and optionally TheTVDB, fanart.tv, or AniDB. That means an outbound request keyed to the title you are adding. You can avoid it by supplying local .nfo metadata and artwork, in which case Jellyfin reads locally and the lookups stop. So the egress here is opt-out and content-specific, a query about a single title, not a stream of your viewing behavior. It is not nothing, but it is small and it is yours to switch off.

What Genuinely Stays Local (the Part That Matters Most)?

The content, which is the whole point. On a self-hosted setup, file and document contents, photo and video libraries, message bodies, and AI prompts and outputs stay on your hardware, and so does the bulk of behavioral telemetry. The continuous stream describing what you opened and when, the thing cloud SaaS collects as its product, is simply not generated for a third party.

It would be dishonest to catalog the leaks without being equally precise about what stops leaving, since that is the substance of the self-hosting case.

• File and document contents. Files in a self-hosted Nextcloud or Seafile live on your disk. The sync payload moves between your devices and your server, not through a provider that can read, index, or train on it.
• Photo and video libraries. A self-hosted Immich keeps the actual images and their analysis, the faces and places, on your hardware. There is no cloud copy being scanned.
• Message bodies. On a self-hosted chat server, message content stays on your server, and with end-to-end encryption it is unreadable even to the server admin. (As the last section noted, federation metadata is the exception, not the message content.)
• AI prompts and outputs. When a model runs on your own machine, your prompts are processed locally and never transmitted to a third-party AI service. This is the cleanest "no pipe" case, because the inference itself happens on-device.
• The bulk of behavioral telemetry. The continuous record of what you opened, when, for how long, and in what order is simply not generated for a third party, because there is no third party in the loop for the core app function.

The structural contrast is the takeaway. In the cloud model, your content is the input to someone else's system, and your privacy depends on that company's policies and incentives. In the self-hosted model, your content is the input to your own system, and the only things that leave are the connection-level residues from the earlier sections, most of which you can shrink further.

Which Remaining Leaks Are Choices, Not Trust?

Nearly all of them. Each residual leak has an architectural answer that does not require trusting a privacy policy: run your own encrypted DNS resolver, enable ECH where supported, use ID-only or UnifiedPush for notifications, connect your devices with WireGuard instead of a hosted tunnel, and supply local metadata to stop title lookups. In the cloud you reduce exposure by reading a policy and hoping; self-hosted, you reduce it by changing the wiring.

The reason this whole topic resolves to privacy-as-architecture is that the residues each have a lever, and pulling the lever is a structural change, not a promise you are asked to believe.

The pattern is the entire argument. In the cloud, you reduce exposure by reading a policy and hoping the company means it. Self-hosted, you reduce exposure by changing the wiring, and then you can confirm it with a packet capture on your own network. That is the difference between a privacy promise and privacy by structure.

This is where the Companion Intelligence approach lives. On a Companion Core, the AI and the apps run on hardware you own, so the content (your prompts, files, photos, messages) has no pipe to leave through. That guarantee is architectural, not a policy you are asked to trust. The connection-level residues this post documents are the honest remainder, and CI leans on the same levers rather than a promise: a self-hosted resolver with encrypted DNS, Tailscale for the owner's private remote access, and a Cloudflare Gateway that provisions per-app HTTPS addresses for reaching your apps. If you want the longer version of why cloud convenience quietly compounds its costs, we wrote about the hidden costs of cloud convenience, and on doing this yourself, running your own AI without the cloud covers the adoption side.

The honest line is the one worth keeping. Self-hosting does not make you invisible on the internet. It makes the part that describes you stay home, and it turns the rest into decisions you can inspect. If that trade is the one you have been trying to think through, the people comparing real setups, real DNS logs, and real push configurations are in the Discord, and that is the room where this stops being theory.

Frequently Asked Questions

Works Cited

• Why your DNS traffic is a privacy leak and how to fix it once and for all — How-To Geek
• Encrypted DNS Factsheet — Internet Society
• What Is Encrypted DNS? DoH vs DoT Explained — CleanBrowsing
• OCSP, CRL and Revoked SSL Certificates — DigiCert
• What Is OCSP? How Certificate Revocation Works — SecureW2
• Pi-hole Setup Guide: Block Trackers at Home — Smart Home Perfected
• Best Pi-hole Blocklists (and How to Install Them) — Burst Bytes
• feat: move version checks to our own infrastructure (PR #27450) — immich-app/immich
• feat: version check endpoint (PR #18572) — immich-app/immich
• Immich, High performance self-hosted photo and video management
• Add way to send telemetry (Issue #35619) — nextcloud/server
• Restrict outgoing connections (Issue #38728) — nextcloud/server
• Good-bye ESNI, hello ECH! — Cloudflare
• Encrypted Client Hello: Closing the SNI Metadata Gap — Center for Democracy and Technology
• Security/Encrypted Client Hello — MozillaWiki
• RFC 8744: Issues and Requirements for SNI Encryption in TLS — RFC Editor
• Encrypted Client Hello Approved for Publication — Feisty Duck
• Encrypted Client Hello in Censorship Circumvention (FOCI 2025) — PET Symposium
• Encrypted Client Hello didn't solve censorship — AdGuard DNS
• What encrypted SNI is and why it matters — ExpressVPN
• Host your own push proxy service — Mattermost documentation
• Mobile push notification service — Zulip documentation
• Nextcloud's push notifications for iOS and Android — Nextcloud
• UnifiedPush: a decentralized, open-source push notification protocol — UnifiedPush
• UnifiedPush announcement — F-Droid
• From your phone (UnifiedPush) — ntfy documentation
• Element X still requires fcm.googleapis.com for UnifiedPush (Issue #6551) — element-hq/element-x-android
• Beyond Signal: Regaining Control with a Self-Hosted Matrix Server — Meister Security
• Why Matrix Fails EU Data Privacy Standards — Wire
• Metadata — Jellyfin documentation
• Metadata Management — Jellyfin (DeepWiki)
• Privacy — Nabu Casa
• Remote access, Security aspects — Nabu Casa Support
• I Built My Own VPN with WireGuard and Headscale — SelfHostedApp
• Tailscale vs WireGuard: Which VPN for Your Homelab? — Selfhosted Guides

How do I keep my family's data private without needing to know how to code?

In 2026 a non-technical family can move their photos, files, and passwords off the cloud and onto their own hardware without writing a single line of code. Apps like Immich (photos), Nextcloud (files), and Vaultwarden (passwords) install with one click. The honest version: no coding is required, but it is not zero effort. You need a device that runs all the time and a one-time setup you can do from a web browser. Companion Hub is the one path that never asks you to open a terminal at all.

Picture your family's digital life a year from now. The photos from every phone in the house back up automatically, the password vault everyone shares lives on a small box on a shelf rather than a company's servers, and the files, calendars, and contacts sync across everyone's devices. All of it sits on hardware you own.

That is the destination, and in 2026 a family can reach it without anyone in the house learning to code. One honest note up front, because you deserve to know what you are signing up for: no coding is not the same as no effort. You will need a device that runs all the time, and you will follow a setup guide once, from a web browser, the way you would set up a new router. The claim is no code, not no learning, and the learning curve in 2026 is far lower than it was a few years ago.

This post is the map: what your private setup looks like, what family data is most exposed right now, which apps replace the cloud services you already use, what "no coding" means in practice, and the one path that never asks you to open a terminal.

What Does Your Family's Digital Life Look Like When It Runs on Your Own Hardware?

On your own hardware, your family's photos back up automatically from every phone, the same way they do to Google Photos today, except they never leave your home. A shared password vault works on every device. Files, calendars, and contacts sync for everyone. The data lives on a small box in your house instead of a company's data center.

You are not giving anything up. The photo timeline still scrolls the way you expect, the passwords still autofill in your browser, the shared folder still holds the tax documents and the schoolwork. What changes is the address: your family's most personal data lives on a device you can point to, governed by your decisions rather than terms you did not write. A capable home setup like this used to require either a cloud subscription or a part-time hobby in system administration. In 2026 there is a third option, finally built for people who just want it to work.

What Family Data Is Most Exposed on the Cloud Right Now?

The most exposed family data lives in a few cloud services: photos and videos, file storage, the password manager, contacts, and viewing history. The common thread is that this data is intimate, kept for years, and governed by a vendor's terms rather than your family's choices. The LastPass breach (2022, disclosed 2023), in which encrypted vaults and metadata were taken, is the clearest reason families move their password vault onto hardware they control.

None of this is meant to alarm you. Most of these services are useful, and most days they work fine. The point is narrower: this is intimate data, kept for years, governed by a company's terms rather than your family's decisions. When the rules change, you find out after the fact.

The password vault is the sharpest example, which is why it is usually the first thing families move. After the LastPass breach, the case for keeping your vault on hardware you control stopped being abstract. The next section names the tool that does it.

The shift is not about hiding. It is about deciding where your family's most personal data lives, and who can reach it.

Which Apps Replace the Cloud Services Your Family Already Uses?

The cloud apps most families use have mature self-hosted replacements in 2026. Immich replaces Google Photos with automatic phone backup, face detection, and shared albums. Nextcloud replaces Google Drive or iCloud for files, calendar, and contacts, with a private account for each family member. Vaultwarden replaces LastPass and stores nothing outside your hardware. Each installs with one click.

Start with photos, because for most families that is the big one. Immich is the closest thing to a self-hosted Google Photos in 2026. It was built from day one to feel like it: your family's photos back up automatically, from every phone, to your own hardware, except they never leave your home. Timeline browsing, face detection, content-based search, shared albums, and a separate account for each person all work. If you have ever been quietly uneasy that every photo of your kids uploads to a system you do not control, this is the change that fixes it.

Two notes on this table. Nextcloud gives each family member their own account with private storage plus shared folders, and calendar and contact sync work on iPhone and Android out of the box. The two streaming rows are where self-hosting stops short, on purpose: Navidrome and Jellyfin are wonderful for music and media you already own, but they cannot replace the Spotify or Netflix catalogs.

What Does "No Coding" Actually Mean?

"No coding" means you need no programming knowledge, no command line (with Companion Hub), and no server administration skills. What you do need is a device that runs all the time and the willingness to follow a setup guide once, from a web browser. The claim is no coding, not no learning. Here is the split.

What you do not need: programming knowledge, command-line experience (with Companion Hub you never see one), knowledge of the plumbing developers use to run these apps, or any server administration.

What you do need: a device that runs all the time, a connection to your home network, the willingness to follow a setup guide once (Companion Hub keeps this to plug-in and a few browser steps), comfort managing user accounts in a web page much like a router admin page, and the occasional software update, which means clicking "update" in a dashboard.

And where friction can appear, because pretending it cannot would not be fair to you:

• Reaching your server from outside the house needs either a simple app called Tailscale or a router setting called port forwarding. Both have non-technical guides, and Tailscale is the easier of the two.
• If your hardware fails, recovery depends on whether you set up backups. Owning your data also means owning the backup.
• Not every cloud feature has a perfect match, which is the subject of the last section.

The learning curve in 2026 is far lower than it was in 2023, and with the right setup the path is consumer-grade.

What Is the One Path That Never Asks You to Open a Terminal?

CasaOS and Umbrel are good non-technical options, but each still assumes you will look at a terminal at least once, such as the one-line command that installs CasaOS. Companion Hub is the one path that removes that step entirely: no terminal ever, no install command to paste, no container files to manage, and a consumer interface that comes pre-configured on Companion Core hardware.

Let me be fair to the alternatives first, because they are good. CasaOS installs with one line and then runs entirely from a friendly web dashboard, with one-click apps including Immich, Nextcloud, and Vaultwarden. Umbrel has the cleanest, most curated feel of any home-server software, especially on dedicated hardware. For a family willing to look at a terminal exactly once, both are solid.

The catch is in that phrase: exactly once. For most people that step is fine. For a parent who has never opened a terminal and does not want to start, it is a wall.

Companion Hub removes the wall. The difference is specific: no terminal ever, with no command to copy and paste, not even once. No knowledge of the developer plumbing (the containers and configuration files) that runs underneath. A consumer interface designed for first-time self-hosters. And it comes pre-configured on Companion Core hardware, so you plug it in, connect it, and go.

That is the concrete answer to this post's question for a reader who has never used a terminal and never wants to. Hub is the one path that asks nothing of the command line at all, because the interface was finally designed for families rather than for engineers.

What Hardware Do You Need to Run This?

Self-hosting needs one thing: a device that stays on all the time, like a small appliance in a closet. For most families the choice is between a Companion Core (pre-configured, around $3,600) or a spare mini-PC running CasaOS (about $150 to $300) for a family comfortable with one terminal step at install.

Two paths fit most families:

• Companion Core. Pre-configured, no setup, about 20W at idle (roughly a couple of LED bulbs). It runs the photos, files, and passwords basics, and adds private local AI on top: AI search across your own files, and Companion Memory. Around $3,600. Choose this when you want it to simply work, and when you also want private AI in the house.
• A spare mini-PC running CasaOS. Around $150 to $300, or free if you already have a spare machine. Good for private photos, files, and passwords, as long as someone is comfortable with one terminal step during the install. Not suitable for local AI.

If the only goal is private photos, files, and passwords, the cheap mini-PC is a real option. Companion Core makes sense when the family also wants local AI, and when nobody wants to touch a terminal at all.

For the curious, here is the fuller set of options with the trade-offs named. You do not need this table to get started.

What Does Self-Hosting Not Solve?

Self-hosting does not make your family invisible. Some apps still call outside services for specific features, your phone's operating system still reports to Apple or Google, and there is no self-hosted replacement for the Netflix or Spotify catalogs. What it does solve is where your photos and files live and who can reach them, a password vault that stays on your hardware, and your family's contacts and calendar.

A setup that oversells itself is not one you can trust, so here are the limits in full.

What self-hosting does solve is the part that matters most to a family: where your photos and files are stored and who can reach them, your password vault (Vaultwarden stores nothing outside your hardware), your family's contact list and calendar, and your media library for the content you own.

Self-hosting gives you sovereignty over the data layers you control. It does not make your family invisible to every digital system, and anyone who promises that is selling something. What it gives you is a meaningful, achievable reduction in cloud exposure, with no coding knowledge required.

If this is the setup you have been wanting for your family, Companion Hub is where you can see what one-click, no-terminal self-hosting looks like. You do not have to decide anything today. You can just look.

Frequently Asked Questions

Works Cited

• CasaOS vs Umbrel vs YunoHost 2026 (Pi Stack)
• CasaOS Review 2026 (MakerStack)
• Cosmos vs CasaOS vs Umbrel (SumGuy)
• Umbrel Review 2026 (BlockDyor)
• Best Self-Hosted Photo Manager 2026 (HomeNode)
• Jellyfin vs Immich vs Nextcloud Photos 2026 (JellyWatch)
• Nextcloud vs Immich (CloudBasedBackup)
• The Complete Self-Hosted Productivity Stack 2026 (CodeROasis)
• Private Cloud and AI on Windows: Nextcloud, Immich, Tailscale (Medium)
• Companion Intelligence Hardware (ci.computer/hardware)
• Companion Intelligence (ci.computer)

How does an AI assistant remember you between sessions without storing history in the cloud?

It uses the same three steps every memory system uses: extract facts from your conversations, store them as vector embeddings in a database, then retrieve the relevant ones and inject them into the model before it answers. The only difference is location. A local system runs all three steps on your own hardware, so nothing leaves your network.

There is a small, specific moment that makes people look at their AI a little differently. You open a fresh chat, ask about something new, and the assistant already knows you prefer short answers, or it remembers the side project you mentioned three weeks ago. You never re-typed any of it. It just knew.

The first reaction is usually delight. The second, if you slow down, is quieter: where is that kept? Something about you is now stored somewhere, and you did not decide where. This post is about how that memory works, step by step, and the one variable that changes everything once you see the mechanism. The mechanism is not the interesting part. The location is.

What Does It Mean for an AI to Remember You?

Remembering you means three things under the hood. First, extraction: an LLM pulls out facts worth keeping, like "prefers concise answers" or "is building a Shopify store." Second, storage: each fact becomes a vector embedding in a database. Third, injection: at the start of a new session, the most relevant memories are retrieved and placed in the model's context before you type a word.

Each step sounds technical and is not. Extraction is note-taking: the model reads what was said and writes down the facts likely to matter later. These are atomic memory units, small discrete facts, not whole transcripts. Your entire chat is not the memory; a handful of distilled facts are. Storage is where the word "vector" shows up, and it is plainer than it sounds: each fact becomes an embedding, a numerical fingerprint of its meaning, filed in a vector database next to the original text. When the system wants what is relevant, it does not match keywords. It searches by meaning. Injection is the payoff: at the start of a session, that similarity search grabs the most relevant memories and slides them into the model's context window before you begin. The model knows things about you because it was just handed them.

Here is the part worth sitting with. This process, extract then store then retrieve and inject, is the same whether it runs on a company's servers or on a machine in your house. The architecture is not proprietary. The only real variable is whose hardware runs it, and who controls the store.

How ChatGPT's Dreaming V3 Remembers You (and Where That Lives)

On June 4, 2026, OpenAI launched Dreaming V3, a background process that reads across your entire conversation history and builds a persistent memory profile. That profile is not stored inside your chat log. It lives in a separate data layer on OpenAI's servers and is injected into the system prompt at inference time, before you type anything.

This is a real step up from the older approach. The previous "saved memories" system would explicitly tell you what it was holding onto. Dreaming V3 synthesizes implicitly, reading patterns across all your sessions without necessarily surfacing what it captured. It weighs three things: freshness (recent events count for more), continuity (how your preferences evolve), and relevance (the right memory for the moment).

The detail that matters most is where the memory sits. It is not buried in your conversation history. It is maintained as a separate data layer on OpenAI's servers, then injected into the model at inference. That has a consequence most people miss: deleting a conversation does not delete the memories extracted from it. Those persist in the memory layer and must be cleared manually through Settings. On free and Plus plans, the conversations themselves are stored indefinitely unless you delete them too.

You do get controls. You can view, edit, and delete saved memories. Temporary Chat keeps a conversation out of your history and out of training data, though it still lives on OpenAI's servers for up to 30 days before deletion. ChatGPT Enterprise offers a full opt-out from training-data use. The point is not that controls are missing. It is that the memory of you lives on infrastructure you do not operate.

What That Means for Your Privacy

A behavioral profile of you, built from your history, preferences, and stated life circumstances, is maintained on a third-party server and injected into every new conversation. You can view and edit a list of saved memories, but a synthesized profile is not the same as a simple editable list. This is not a claim of wrongdoing. It is a description of where the data sits.

A 2026 study presented at the ACM CHI Conference named the tension well, and the plain version is the one that lands: the feature you value most is also the one you cannot fully see or constrain. The editable list shows the facts the system chose to surface. But Dreaming V3 reads patterns across sessions, and those patterns may never appear in any readable form. You can edit the list. You cannot necessarily see the synthesis.

The law has started to name this too. Under the EU's GDPR, AI systems that build persistent behavioral profiles are classified as profiling activities, which triggers consent obligations and the right to erasure. The EU AI Act's transparency obligations for chatbot systems take effect August 2, 2026, less than two months after Dreaming V3 shipped. The direction is clear: a profile of you, held somewhere else, is something you are owed visibility into.

How Local AI Memory Works Without the Cloud

Local memory uses the identical architecture, but every component runs on your machine. A local vector database such as Chroma, Qdrant, or pgvector stores the embeddings. Local embedding tools convert text to vectors on-device with no external API call. At session start, a similarity search retrieves the relevant memories and injects them into the model. Nothing is transmitted off your network.

Walk it through the same three steps and notice nothing is missing. The vector database runs on your own hardware and holds the text plus its embeddings. The embedding step runs on-device too: a tool like Mem0 with FastEmbed converts conversation text into vectors without sending anything out. And retrieval-augmented generation, usually shortened to RAG, does the injection, pulling the relevant memories from your local store into the context window before the model responds.

The 2026 state of the art goes further with hybrid memory, combining three kinds of recall: vector search for unstructured, intuition-style retrieval; graph traversal for structured facts and relationships; and episodic storage for temporal sequences, the "last Tuesday she mentioned the client deadline" kind of memory. Together they give a local system real depth. This is not a thought experiment. Mem0 has published a working guide for building exactly this on a local stack using Mem0, OpenClaw, and Ollama, the same kind of stack Companion Intelligence builds on. The local memory layer is not a promise; it is a documented, running pattern.

What Today's Local Tools Offer (the Honest Gap)

Most local tools were built for private chat, not persistent personal memory. AnythingLLM has built-in retrieval but keeps memory per workspace rather than unified. LM Studio starts every session blank unless you configure it. Jan.ai is fully offline but is not designed for cross-session memory. Mem0 solves the memory layer technically; Companion Memory solves it as a finished consumer product.

There is a gap worth stating plainly, because pretending it does not exist would be the wrong kind of pitch. The memory layer is a separate engineering problem from running a model locally, and not every tool has solved it. The deeper point is about quality. Without a continuous background synthesis process like Dreaming V3, local memory leans on explicit extraction: it stores what you tell it to remember, or what a simple extraction model catches in conversation. Right now, cloud memory is more sophisticated in synthesis depth. Local memory is architecturally equivalent in its core mechanism but usually asks for more intentional setup. Mem0's 2026 hybrid architecture narrows that gap meaningfully; it does not erase it out of the box.

The Real Question Is Where the Vector Store Lives

The memory question and the privacy question collapse into one: where does the vector store live? If it is on OpenAI's servers, that is their infrastructure, their governance, their terms of service. If it is on your hardware, the memory is equally capable (extraction, embedding, retrieval, injection) but none of it leaves your network. That is an architectural difference, not a philosophical one.

This is what Companion Memory is built to do. It implements the full memory architecture, extraction, embedding, vector storage, and retrieval-augmented injection, entirely on Companion Core hardware. Your preferences, your history, your context, your professional patterns live on a machine in your home or office. They are not transmitted, not synthesized on a remote server, not governed by a third party's terms of service. The specific technical answer to the question in the title is exactly this: a local vector database and RAG, not a remote synthesis server.

To keep the comparison honest, here is the whole trade-off in one view.

The conclusion is the honest one. Cloud memory is currently more sophisticated in synthesis depth. Local memory is architecturally equivalent in its core mechanism, and it asks for more intentional setup, unless you run a purpose-built product like Companion Memory, in which case the setup cost goes away and the data stays home. For more on how cloud convenience quietly compounds its costs, we wrote about the hidden costs of cloud convenience, and on whether you can do this yourself, running your own AI without the cloud covers the adoption side directly.

Once you see the mechanism clearly, the decision is not about whether an AI can remember you without the cloud. It plainly can. The decision is whose machine you want that memory to live on. If you would rather it live on yours, that is what a Companion Core, with Companion Memory underneath it, is for. If that is the version of this you have been waiting for, it is worth a look at the store.

Frequently Asked Questions

Works Cited

• Dreaming: Better memory for a more helpful ChatGPT — OpenAI
• ChatGPT Memory Dreaming Update — TechTimes
• ChatGPT Dreaming V3: New memory architecture — Windows News
• Memory FAQ — OpenAI Help Center
• Is ChatGPT Safe? — AskSafely
• State of AI Agent Memory 2026 — Mem0
• The AI Memory Layer Guide — Mem0
• Adding Persistent Memory to Local AI Agents (Mem0 + OpenClaw + Ollama) — Mem0
• Mem0: Production-Ready AI Agents — arXiv
• Build a Local AI With Persistent Memory: LM Studio + Big RAG — n1n.ai
• Run a Fully Local AI With Persistent Memory — DEV Community
• AnythingLLM 2026 Setup Guide — Local AI Master
• Is ChatGPT Private? 2026 Guide — NordVPN

Why are 77% of workers leaking company data through AI tools?

According to the LayerX Enterprise AI and SaaS Data Security Report 2025, 77% of employees paste company information into AI tools, and 82% of those paste events come from personal accounts with no enterprise privacy protections. The behavior is structural, not careless: sanctioned tools are slower, most workers lack managed accounts, and traditional data loss prevention cannot see copy-paste into a browser.

I have pasted things into ChatGPT I should not have. A client's contract language, to rewrite it faster. A block of our own code, to find a bug before a call. I was not being reckless. I was trying to finish something, and the tool in the browser tab was faster than the careful path. If you run a team, several people around you probably did the same thing this week, and none of them think of it as a leak.

That is the honest starting point for the number everyone keeps quoting. So before we treat 77% as a scandal, here is what it actually measures, why the behavior keeps happening, and the one change that removes the leak instead of just scolding the people doing it.

What does the 77% statistic actually mean?

The 77% figure comes from the LayerX Enterprise AI and SaaS Data Security Report 2025, covered during Data Privacy Week in January 2026. It is behavioral data captured from real enterprise browser activity, not an opinion survey. It measures employees who paste company information into AI tools, and 82% of those events come from personal accounts.

That distinction matters. This is not a poll where people admitted they might paste something sensitive someday. It is a record of what actually moved through the browser. The same report found employees average 14 paste events per day through personal accounts, and that copy-paste into AI has become the single largest channel for corporate data leaving company control. The number is large because the behavior is ordinary, repeated, and quiet.

So the right way to read 77% is not "a crisis just landed." It is a measurement of scale on something already happening everywhere, finally counted. The interesting question is not how shocking the figure is. It is why it is so high.

This is not carelessness. It is structure.

Most of the 77% are trying to be productive, not leak data. Three structural gaps explain it: sanctioned AI tools are slower and less capable, most workers have no managed enterprise account so 82% use personal ones, and legacy data loss prevention was built for file transfers, not copy-paste into a browser.

The productivity gap comes first. Where an approved AI tool exists at all, it tends to lag the consumer version: slower to get new capabilities, constrained by procurement, often a step behind on quality. An engineer who needs to debug code in the next hour reaches for the tool that works now. That is not a discipline failure. It is someone doing their job with the better instrument.

The account gap follows. Those 82% are on personal accounts mostly because nobody handed them a managed one. The employer never bought enterprise seats, never configured them, or never told anyone they existed. A personal account offers none of the enterprise privacy protections, but it is the only door that is open.

The third gap is the one IT teams feel most sharply. Traditional data loss prevention was designed to watch file transfers, email attachments, and USB drives. It was never designed to watch someone copy a paragraph and paste it into a browser tab. Browser-based AI prompts are functionally invisible to most enterprise DLP. The tooling meant to catch data leaving the building cannot see this door at all.

What goes through that door is what you would guess: proprietary source code, client contracts and correspondence, internal meeting transcripts. The pattern is the useful part. The most valuable, least-shareable material a company holds is exactly the material people most want an AI to help them with, which is why it ends up in the prompt box.

What happens to data you paste into a personal ChatGPT account?

On non-Enterprise ChatGPT tiers, OpenAI may use your conversations to train its models by default, and once content is submitted it cannot be recalled from the training pipeline. ChatGPT Enterprise differs: training opt-out is on by default with audit logs. But 82% of corporate paste events run through personal accounts without those protections.

This is the mechanism most people have never been walked through. On the free and Plus tiers, OpenAI's default terms allow submitted conversations to be used to improve its models. You can turn that setting off, but the default is opt-in, and the data you already submitted does not come back. There is no recall button for a prompt that has entered a training pipeline.

ChatGPT Enterprise is a different posture: training opt-out on by default, data not used to improve the models, zero data retention available through the API, company-managed accounts with audit logs. If every paste event went through a properly configured Enterprise account, this article would be much shorter.

The gap is the whole story. 82% of corporate paste events go through personal accounts with none of those protections. The person who pasted source code into their own ChatGPT account has, under the default terms, potentially handed that code to model training. Not because they ignored a warning, but because no one ever showed them the difference between the two doors.

What this looks like when nothing stops it: the Samsung pattern

Within 20 days of allowing ChatGPT, Samsung's semiconductor division had three data-exposure incidents: proprietary source code pasted twice, and a recorded meeting transcribed and pasted for minutes. Each was a reasonable attempt to work faster, not a mistake to mock. It shows what default behavior looks like when no structural constraint exists.

Here is what happened, in April 2023. An engineer pasted faulty source code from a semiconductor measurement database into ChatGPT to find a fix. A second engineer pasted optimization code for spotting defective equipment. A third recorded a meeting, converted it to text, and pasted it in to generate minutes. Three incidents, twenty days, one division.

It is easy to read that as a story about careless engineers, and I would ask you to resist it. Every one of those actions is something I would have been tempted to do, and probably something you would too. Debug this faster. Summarize this meeting for me. None of it is negligent on its face. It is what capable people do with a powerful tool and a deadline when nothing in the system stops the data from leaving.

That is the value of the Samsung case. It is not a cautionary tale about other people's bad judgment. It is a clean demonstration of the default: give competent employees a useful AI and no structural constraint, and the most sensitive material in the company starts flowing outward, not through malice, but through ordinary productivity. Samsung's response, a company-wide ban plus a 1,024-byte limit on prompt length so tight it made the tool nearly useless for code work, brings us to the fix that does not fix anything.

Why the obvious fixes do not work

Banning the tool or writing a no-paste policy does not solve the productivity problem; it relocates the behavior to another tool. A policy is unenforceable at the session level because data loss prevention cannot read browser prompt inputs. An enterprise account improves the terms but keeps a governance dependency on a third party.

"Do not paste confidential data into AI tools" is a sentence in a handbook. It is unenforceable at the level where it matters, the individual session, because the DLP layer cannot see browser prompts. A rule no system can observe depends entirely on every person remembering it every time they are busy. The 77% number is what that dependency produces.

The ban fails differently. Blocking the tool does not remove the reason people reached for it, so the behavior moves: to a personal phone, a home laptop, whatever is not yet blocked. You have not closed the door, you have moved it somewhere you can no longer see. The research is candid that employee training and tool bans are known non-solutions, not because awareness is worthless, but because neither one changes the structure that produces the leak.

An enterprise account is the most honest partial fix. It improves the terms. But notice what it does not change: your most sensitive data still leaves your network and lands on someone else's servers. You have upgraded the contract that governs what they may do with it. You have not removed the dependency on that contract being honored, or on that company never being breached or compelled. The data is still out there. You are now trusting a better promise.

The structural fix: move the AI onto hardware you control

The only way to guarantee company data never reaches a third-party AI is to run the AI in an environment you control. The guarantee is architectural, not contractual: data cannot leave because there is no pipe. Local models via Companion Core and Hub apps process prompts on your own hardware, without transmitting anything outward.

Every fix in the previous section tries to govern data that is leaving the building. The structural answer is to stop the data from leaving at all. Not by watching the door harder, but by not having that particular door.

When the AI runs on your own hardware, the sequence is simple. Your prompt goes to a model on a machine you own, which processes it and returns an answer. Nothing is transmitted to an outside server, because there is no outside server in the loop. A policy promise says "we will not misuse your data." An architecture says "your data did not go anywhere it could be misused." Those differ in kind, not degree.

This is what Companion Intelligence builds. Local AI models through Companion Core process your prompts on hardware in your own space. Hub apps run the same way: Taiga for project management, Nextcloud for files, Plane for issue tracking, all on infrastructure your organization controls. The tasks people reach for an AI to do, analyze a document, review some code, summarize a meeting, draft a first pass, are exactly the tasks that can run locally. The point is not to take the useful tool away. It is to put it somewhere the data never has to leave.

Honest limits

Running AI locally removes the leak vector, but it is not free or frictionless. Local models on modest hardware trail the best cloud models on the hardest reasoning tasks, and self-hosting carries a setup cost that turnkey options reduce but do not erase. The 77% figure also has a freshness window, though the structural point holds regardless.

On capability: for the everyday work that fills most days, writing, summarizing, document analysis, ordinary coding help, a competent local model is hard to distinguish from the cloud. For the hardest edge of reasoning, the frontier cloud models still lead. If your work lives almost entirely at that frontier, weigh that trade honestly.

On effort: a self-hosted setup is not zero work. The DIY path takes real time to stand up. Turnkey options exist specifically to shrink that cost, and they shrink it a lot, but "less setup" is not "no setup."

On the number: the 77% figure is a point-in-time reading from January 2026 Data Privacy Week reporting, and it carries a freshness window. The Cyberhaven 2026 AI Adoption and Risk Report corroborates the same pattern with a different methodology, which is reassuring, but treat the exact percentage as a snapshot. The structural conclusion does not depend on the decimal. Whether it is 77% this quarter or a neighboring figure next quarter, the mechanism is the same and the fix is the same.

The question worth changing

If you take one thing from the number, let it be a change of question. The reflex is to ask, "how do I get my team to stop pasting sensitive data into AI tools?" That question has no good answer, because it asks people to win a fight against their own productivity, every day, with no system helping them.

The better question is, "where does our AI run?" That one has an answer you can implement, and it resolves the leak for everyone at once instead of asking each person to resolve it a hundred times a day.

If that reframing lands, the CI Discord is where people are working through what running AI on their own hardware looks like in practice. Come ask the awkward questions. It is the room where this is being figured out in the open.

Frequently Asked Questions

Works Cited

• Data Privacy Week 2026: Why 77% of Employees Are Leaking Corporate Data (Breached.Company)
• LayerX Enterprise AI and SaaS Data Security Report 2025
• AI Is Now the Number One Data Exfiltration Vector (LayerX)
• 77% of Employees Leak Data via ChatGPT (eSecurity Planet)
• Cyberhaven 2026 AI Adoption and Risk Report
• Samsung Employees Leaked Corporate Data in ChatGPT (CIO Dive)
• Samsung Engineers Feed Sensitive Data to ChatGPT (Dark Reading)
• AI Incident Database, Samsung ChatGPT Incident 768
• OpenAI Data Controls FAQ
• OpenAI Enterprise Privacy
• 89% of Enterprise AI Usage Invisible to the Organization (Help Net Security)

I have spent years watching the terms of a tool change after I had already built my work around it. Not a price bump you can shrug off, but the kind where the thing you depend on quietly starts doing something new with what you put into it. That is where a lot of teams are right now with Jira and Confluence, and the clock is real: the policy takes effect August 17, 2026.

This is the calm version of that news. What the policy says, what data it touches, how to turn off the part that matters in about five minutes, how Atlassian's choice compares to a few of its peers, and the deeper question underneath it. No alarm. Just the facts and what you can do with them.

What Atlassian's New Policy Actually Says

Atlassian announced in April 2026 that, effective August 17, 2026, it will use customer data from Jira, Confluence, and Jira Service Management to train its AI products by default. The change affects roughly 300,000 cloud customers. Two categories are covered: de-identified metadata and in-app content. Collected data may be retained for up to seven years. The two categories are worth separating, because the opt-out story is different for each.

• Metadata. De-identified, aggregated signals: readability and complexity metrics, task taxonomies, semantic similarity indices, iteration completion timestamps, Jira SLA values. Atlassian de-identifies and aggregates this before it reaches a model.
• In-app data. User-generated content: Confluence page titles and body text, Jira issue headers, descriptions, comments, and custom names for emojis, statuses, and workflows. This is the category that contains your team's actual confidential work. Metadata is the abstracted shape of how your team works. In-app data is the work itself. When people worry about "Jira training AI on our data," the in-app category is the one they mean, and it is the one you can do something about on every plan.

What Data Is Affected, and Who Can Opt Out

Every plan can opt out of in-app data, the actual content your team writes. No plan below Enterprise can opt out of metadata collection. Enterprise customers have both off by default. Full exemptions apply to customers using customer-managed encryption keys, Atlassian Government Cloud, Atlassian Isolated Cloud, and HIPAA-covered customers. Here is the operational picture by subscription tier. This is the table to screenshot.

A few takeaways fall out of that grid. Every plan can opt out of in-app data, the critical category, so that is the action that matters most. No plan below Enterprise can opt out of metadata. Enterprise is protected by default on both counts. Standard needs the most attention: in-app data is opt-out-required, and metadata cannot be turned off at all.

There are also complete exemptions, where no data is collected at all: customers using customer-managed encryption keys, Atlassian Government Cloud customers, Atlassian Isolated Cloud customers, and HIPAA-covered customers. If you are not in one of those buckets, the defaults above apply to you.

How to Turn It Off Before August 17 (5 Minutes)

Sign in to your Atlassian organization at admin.atlassian.com, go to Security, then Data contribution, and turn off "Contribute in-app data for AI model training." Save. The change applies to your whole organization. If you run multiple Atlassian sites, repeat for each. The setting has been live to configure since April 16, 2026.

• Log in to your Atlassian organization at admin.atlassian.com.
• Go to Security, then Data contribution.
• Turn off "Contribute in-app data for AI model training."
• Save. The change applies to your entire organization.
• If you have multiple Atlassian sites, repeat for each site. That is the whole procedure. No contract changes, no sales call, about five minutes. The only constraint is timing: configure it before August 17, 2026. The toggle is already there waiting for you. If you administer more than one site, the per-site repeat is the easy thing to forget, so it is worth a second pass.

How This Compares to Notion, Slack, and Microsoft

Atlassian opts customers in by default. Several peers do not train on customer data by default: Notion states using Notion AI does not grant it the right to train models on your data, Slack does not train generative AI models on customer data without affirmative opt-in, and Microsoft prevents foundation-model training and offers data residency controls.

This is the published policy for each tool, with primary sources linked below. It is not a verdict on any company, just the documented posture each one has chosen.

One nuance is easy to lose. Atlassian's policy applies specifically to its own AI product development, including Atlassian Rovo and Rovo Dev. It is not claiming the data for general commercial sale. The mechanism is that your team's content informs how Jira's AI gets smarter, and "smarter for Atlassian's AI" is built on your team's specific confidential work. That is the concern worth weighing.

What "Training on Your Data" Actually Means

Machine learning training uses existing data to adjust a model's parameters, its pattern-recognition weights, so it produces better outputs later. Training on your content makes your writing patterns and project structures part of the statistical substrate shaping future responses. Atlassian de-identifies and aggregates the data first, which reduces re-identification risk but does not remove the contributed patterns.

Precision is the antidote to panic here. When a model trains on your Confluence pages and Jira issues, your writing style, your issue templates, your team vocabulary, and your project structures become part of the statistical substrate that shapes how the system responds to future users. De-identification lowers re-identification risk, but it is not the same as removal: the patterns your team's work contributes persist in the model's behavior even after your organization's footprint is obscured. Your product roadmaps inform its understanding of how roadmaps are structured; your client project descriptions inform its sense of how service work is described.

This is not a claim that something nefarious happens to a specific document. It is a statement about where the control sits. Once the content is in the training set, the decisions about it are no longer yours.

The Deeper Pattern: Why the Location of Your Data Decides What Can Happen to It

When your team's work runs in cloud software, the vendor sets the terms for what happens to that content, and those terms can change. Self-hosted Jira (Jira Data Center) does not train Atlassian's AI because the data never reaches Atlassian. The location of your data determines what can be done with it, more reliably than any terms-of-service commitment.

The Atlassian situation is a specific instance of a broader pattern. When you run your team's work in cloud software, the conditions of what happens to that content are set by the vendor, not by you, and they can be revised. In April 2026, Atlassian updated its terms. In May 2025, Plex placed remote streaming behind a paywall. In January 2026, Spotify raised prices. Each is a unilateral change to a service you cannot inspect and cannot easily replace in the middle of a workflow.

The architectural answer is not to keep hunting for a better-behaved SaaS vendor. It is to run the infrastructure you depend on from hardware you control. Self-hosted Jira (Jira Data Center) does not train Atlassian's AI on your data, because the data never reaches Atlassian. That is the privacy-as-architecture idea in one sentence: the location of your data determines what can be done with it, more reliably than any promise in a terms-of-service document. A policy can change in April. A structural fact, the data sitting on hardware you own with no pipe leaving it, does not change because a vendor revised a clause.

The honest part: self-hosting is not free, and it is not effortless. Jira Data Center requires server infrastructure and ongoing IT management. Running your own stack means you own the uptime, the backups, and the updates a cloud vendor otherwise handles. That is a real trade-off, and for some teams the cloud convenience is worth the exposure. The point is to make that a decision you reach on purpose, not a default you inherit.

What You Can Do (A Graduated Response)

Immediately, on any plan, opt out of in-app data contribution before August 17. On Standard, where metadata cannot be opted out, teams handling sensitive client work can weigh an Enterprise upgrade or an alternative. Structurally, self-hosted options (Jira Data Center, Plane, GitLab, Taiga, NocoDB) keep project data on hardware you control.

Three levels, depending on how far you want to go. Immediate, all plans, before August 17, 2026. Opt out of in-app data contribution now: Admin, Security, Data contribution, off. Cost: about five minutes, no contract changes.

Medium-term, for Standard plan users. Metadata collection cannot be opted out on Standard. If metadata privacy matters for your team's work, particularly for teams handling sensitive clients, evaluate whether an Enterprise upgrade is warranted or whether alternatives are worth exploring.

Structural, for teams thinking architecturally. If this episode has you reconsidering where your project data lives, these are the self-hostable paths named in our research:

• Jira Data Center, Atlassian's own self-hosted product. Data stays on your servers, with no AI training contribution. Requires server infrastructure and IT management.
• Plane, an open-source, self-hostable project management platform launched on GitHub as a Jira alternative. Full issue tracking, sprints, and modules, deployed on your own hardware.
• GitLab (self-managed), a complete DevOps and project management stack. GitLab's blog explicitly positioned itself as an opt-out alternative when Atlassian's policy broke in April 2026.
• Taiga, available on Companion Hub. Open-source Scrum and kanban project management, one-click deploy, running entirely on your hardware with no cloud AI training.
• NocoDB, available on Companion Hub. A database-and-spreadsheet hybrid that can replace Confluence-style content databases, self-hosted, with no external data contribution. You do not have to leave Jira to protect your team this month. The five-minute opt-out does the urgent work. The structural options are there for when you are ready to ask the longer question.

Frequently Asked Questions

Does opting out of in-app data stop all data collection?

No. Opting out of in-app data stops collection of the content your team writes. On Free, Standard, and Premium plans, de-identified metadata collection continues and cannot be opted out. Only Enterprise has metadata off by default.

What happens if I do nothing before August 17, 2026?

On plans where in-app data is opt-out-required, doing nothing leaves your team's content eligible to be used for Atlassian's AI training when the policy takes effect. You can change the setting later, but content already contributed is the concern, so acting before the date is the safe move.

Is my Jira data sold to other companies?

No. According to the policy, the data is used for Atlassian's own AI product development, including Atlassian Rovo, not sold for general commercial use. The concern is that your confidential content shapes how Atlassian's AI behaves, not that it is resold.

Does self-hosted Jira train Atlassian's AI? No. Jira Data Center, Atlassian's self-hosted product, keeps your data on your own servers, so it never reaches Atlassian and cannot be used for AI training. The trade-off is that self-hosting requires server infrastructure and IT management.

How would I know if another SaaS tool is training AI on my data? Often you do not know until the policy changes, usually through a notification email and a settings page, with the default increasingly set to opt-out-required. The structural resolution is the same: if the data never leaves your network, the question does not arise.

The thing I keep coming back to is that none of this is about Atlassian, not at its core.

Atlassian made a defensible product decision and announced it in advance, which is more than some companies do. The clarifying part is what the moment reveals: when your work lives somewhere you do not control, the rules can change, and the burden to notice and respond lands on you.

So do the five-minute thing today if you are on a cloud plan. Then, if it has you thinking about where your team's work lives, that is the more interesting conversation. If you are working through what this means for your setup, the Companion Intelligence Discord is where people are comparing notes and figuring it out together.

If you have ever done that math, you know the feeling, but is another service to help clean-up the mess what we really need?

The average household pays for eight to twelve cloud app subscriptions. Most people do not pay for all of them, but four or five is common, and that quietly adds up to forty to sixty dollars a month on the low end. Can we really afford this?

There are local options

A few years ago, the honest answer to "can I just run this myself?" was "yes, if I learn Docker and give up a significant amount of time." In 2026 that answer is mostly just "yes." The setup tax that bounced people off self-hosting has largely collapsed, so the real question is not whether you can replace an app. It is which one you start with.

This post is a map, not a sales pitch. It walks through the cloud apps you can replace with a self-hosted server, which of them are already packaged on Companion Hub for one-click install, what you would realistically save, and an honest section on what self-hosting still does not replace well. Bring your own list of subscriptions and check them off as we go.

How Much of Your App Stack Are You Actually Renting?

The average household pays for eight to twelve cloud app subscriptions. Most people pay for four or five without noticing, roughly forty to sixty dollars a month. Stacked together, common tools like storage, notes, a password manager, video calls, and document signing can pass one hundred dollars a month if you pay for all of them.

Here is the stack a lot of people are renting, with monthly costs as of mid-2026:

Pay for all nine and you are at about $108 a month, roughly $1,296 a year, and close to $3,888 over three years. Most people do not pay for the full set. The point is not the maximum. It is how little it takes to cross forty or fifty dollars a month without ever making a deliberate decision to spend it.

What You Can Replace, Category by Category

Nearly every common category has a self-hosted answer in 2026: Nextcloud or Seafile for files, Immich for photos, AppFlowy or AFFiNE for notes, VaultWarden for passwords, Jellyfin for media, Jitsi for video calls, Documenso for signing, and Home Assistant for smart home. Most are packaged on Companion Hub for one-click install. Here is the map, category by category.

• File storage and sync (Google Drive, Dropbox, OneDrive): Nextcloud is the most feature-complete self-hosted cloud (sync, document editing, calendar, contacts). Seafile syncs faster for large libraries. Both install via Coolify on Hub, with no data egress fees.
• Photo backup (Google Photos, iCloud Photos): Immich is the leading self-hosted Google Photos alternative, with mobile auto-upload on iOS and Android, face recognition, shared albums, and a timeline view. Production-ready, free, and on Hub. This is the one most people are surprised actually works.
• Notes, docs, and knowledge base (Notion, Confluence): AppFlowy is an open-source Notion alternative (docs, boards, databases, wikis). AFFiNE covers Notion and Miro in one tool, Excalidraw handles whiteboarding, and Joplin does encrypted markdown notes. AppFlowy, AFFiNE, and Excalidraw are on Hub.
• Password management (1Password, LastPass, cloud Bitwarden): VaultWarden is a Bitwarden-compatible server written in Rust. It runs on Raspberry Pi grade hardware, works with every Bitwarden client, and is the most popular self-hosted password manager in 2026. One-click on Hub.
• Media server, video (Plex Pass, parts of Netflix): Jellyfin is a free, open-source media server with no premium tier and no remote-streaming paywall, unlike Plex, which moved remote streaming behind Plex Pass in 2025. Plex is also on Hub for existing libraries. Both are Hub-available.
• Music streaming (Spotify, Tidal, for owned music): Navidrome streams your owned collection to Subsonic-compatible apps on iOS and Android, and Koel adds a polished interface. Both via Hub. Be clear-eyed here: a self-hosted music server streams the music you already own. It does not replicate Spotify's 100 million track catalog or its discovery. If you have a large local collection the replacement is complete. If you use Spotify mainly to discover, it is not.
• Team chat (Slack, Teams, private Discord): Element runs on Matrix, with end-to-end encrypted chat, rooms, voice and video, federated. Mattermost offers a Slack-identical interface and is used inside Companion Intelligence. Element is on Hub, Mattermost is in the Hub ecosystem.
• Video conferencing (Zoom Pro, Meet): Jitsi Meet is fully open-source, with no time limits and no account needed for participants. Confirmed on Hub.
• Project management and databases (Jira, ClickUp, Airtable, Asana): Taiga does full project management (Scrum, kanban, sprints, issues). NocoDB is a spreadsheet-database hybrid, an Airtable alternative. Both confirmed on Hub.
• AI image generation (Midjourney, Firefly, DALL-E): ComfyUI is a node-based Stable Diffusion interface that runs locally on your GPU, production-ready and confirmed on Hub. The images stay local.
• AI workflows (Zapier, n8n cloud, Make): Flowise is a visual AI workflow builder for agents, RAG pipelines, and automation chains, run locally. Confirmed on Hub.
• Document signing (DocuSign, HelloSign): Documenso is an open-source DocuSign alternative with full e-signature workflows, an audit trail, and PDF support. Confirmed on Hub.
• Smart home control (Google Home, Alexa cloud, SmartThings): Home Assistant runs locally, integrates with more than 3,000 devices, and needs no cloud. Confirmed on Hub. Your home data stays on your home network.

Which of These Are Already on Companion Hub?

As of June 2026, Companion Hub includes one-click installs for AppFlowy, AFFiNE, ComfyUI, Coolify, Documenso, Element, Excalidraw, Flowise, Home Assistant, Jellyfin, Jitsi Meet, NocoDB, PairDrop, Plex, Taiga, and VaultWarden, among others. That packaging is the difference: no Docker, no reverse proxy, no weekend of setup.

This is what separates the list above from a generic self-hosting roundup. Most articles hand you twenty open-source projects and a link to each one's install docs. The hard part is what comes next: spinning up containers, configuring a reverse proxy, sorting out certificates, and hoping the upgrade path does not break. Companion Hub collapses that into browsing a marketplace and clicking install.

The confirmed Hub app list as of June 2026:

• AppFlowy, AFFiNE, Excalidraw (notes, docs, whiteboards)
• VaultWarden (passwords)
• Immich (photos)
• Jellyfin, Plex (media)
• Jitsi Meet (video calls)
• Element (chat), with Mattermost in the Hub ecosystem
• Taiga, NocoDB (project management and databases)
• ComfyUI (AI image generation), Flowise (AI workflows)
• Documenso (document signing)
• Home Assistant (smart home)
• Coolify (the deployment layer that installs Nextcloud, Seafile, and more)
• PairDrop, Keila, and partner apps, with Hermes and OpenClaw available through the Hub ecosystem

One caveat worth stating plainly: this list is current as of June 2026. The Hub store keeps adding apps, so check the live marketplace when you go to install. If something here is not listed, or something new is, the live store is the source of truth.

What Would You Actually Save?

A hobbyist who replaces five common subscriptions (Google One, Notion Plus, 1Password, Zoom Pro, and DocuSign) saves about fifty-five dollars a month, roughly nineteen hundred dollars over three years. That is one example mix. Substitute your own apps to get your number, and it excludes media, AI, and project-management savings.

Here is that example, spelled out:

Over three years, that is about $1,979 on just these five subscriptions, and it does not count media streaming, AI generation, or project management. To be clear, this is one mix chosen to illustrate the math, not a universal claim about your bill. The honest way to use it is to take your own list of subscriptions, find the equivalents above, and run your own total. Your number is the one that matters.

What Self-Hosting Does Not Replace Well

A few cloud services have no clean self-hosted match yet. Team email is doable but complex to run. Self-hosted music servers stream your owned library, not Spotify's catalog or discovery. Real-time document collaboration with outside people, full Adobe Creative Cloud parity, and complex enterprise Jira workflows remain partial replacements rather than drop-in ones.

• Team email. Self-hosted email (with Mailcow, for example) is doable but complex to run and keep deliverable. For critical communication, most people should keep a managed domain-email solution rather than self-host it.
• Spotify's catalog. Navidrome and Koel replace your owned music library well. They do not replicate Spotify's 100 million plus track catalog or its discovery algorithms. If discovery is why you pay for Spotify, a self-hosted server is not a full replacement.
• Real-time collaboration with outside people. Nextcloud and ONLYOFFICE support live document collaboration, but the external person has to connect to your server. It works, with a bit more friction than a shared Google Doc link.
• Adobe Creative Cloud. Krita, Inkscape, and GIMP cover most use cases, and DaVinci Resolve handles professional video editing for free and locally. The gap narrows every year, but for some professional workflows it is not yet zero.
• Jira at scale. Taiga handles most project management well. For complex enterprise Jira workflows, Plane (also self-hostable) is a closer match than Taiga. Pick the tool to the workflow, not the other way around.
• None of this is a reason to stay fully in the cloud. It is a reason to start where the replacement is clean and keep the few hard cases on managed services until the gap closes. Knowing the limits is what makes the rest trustworthy.

Where to Start

Pick one app you already pay for and install its Companion Hub equivalent, then run it for a week before moving the next one. VaultWarden (passwords) and Immich (photos) are common first installs: high value, low risk, and quick to set up. You do not need to migrate everything at once.

That is the whole trick to making this stick. People who try to replace their entire stack in one weekend tend to burn out and roll back. People who replace one app, live with it for a week, and then move to the next one tend to keep going until the monthly bill is a fraction of what it was. Start with the one that annoys you most, or the one that costs you most. Either is a good first move.

If this sounds like your setup, Companion Hub is where the one-click versions of every app above already live. You can browse the marketplace, pick your first replacement, and have it running today.

Frequently Asked Questions

****Do I need to know Docker to self-host these apps?

No. The apps in this post are packaged on Companion Hub for one-click install, so you do not need Docker, a reverse proxy, or command-line setup to run them.

****Can I really replace Google Photos?

Yes. Immich is a production-ready self-hosted Google Photos alternative with mobile auto-upload on iOS and Android, face recognition, shared albums, and a timeline view. It is available on Companion Hub.

****Does self-hosting replace Spotify?

Only partly. Self-hosted music servers like Navidrome and Koel stream music you already own from your server. They do not replicate Spotify's 100 million track catalog or its discovery algorithms.

****How much can a typical person save?

As an example, replacing five common subscriptions (storage, notes, passwords, video calls, document signing) saves about fifty-five dollars a month, near nineteen hundred dollars over three years. Your number depends on the apps you pay for today.

****Are these self-hosted apps actually production-ready in 2026?

Yes for the named categories in this post. The honest-limitations section lists the cases (team email, music discovery, external real-time collaboration, full Adobe parity, complex Jira) that remain partial replacements.

Works Cited

• 10 Self-Hosted Apps You Need to Try, netbird.io
• Best Self Hosted Apps in 2026, Pinggy
• 50+ Open-Source Alternatives to Cloud Services, DreamHost
• I ditched Plex, Google Photos, iCloud for self-hosted alternatives, XDA Developers
• Spotify Premium Price Increase Jan 2026, gamsgo.com
• 1Password Price Hike March 2026, 9to5google
• Notion Pricing 2026, notion.com/pricing
• Companion Intelligence, ci.computer
• Companion Intelligence Hardware, ci.computer/hardware

Is a personal AI server worth it if you are not a developer?

Yes, with honest caveats. For a non-developer entrepreneur in 2026 it is worth it when two things are true: you use AI often enough that the cost math closes, and you handle client data confidential enough that keeping it local matters. The technical barrier has dropped. Tools like LM Studio install in minutes, and Companion Core needs no terminal at all.

The question carries a hidden assumption: that developer skills are the price of entry. In 2024 that was close to true. Setting up local AI meant a terminal, a config file, and a free weekend. In 2026 that assumption is outdated for most of what an entrepreneur needs AI to do.

So this post answers the two questions you are asking, plainly and in order. Can you use this without technical skills? And will it help your business? It also does the thing most local-AI content skips: it tells you where a personal AI server is not worth it, so you can make the call for yourself.

What Can a Non-Developer Do With a Personal AI Server?

A non-developer can run document Q&A, draft client proposals and emails, summarize meeting recordings, review scripts before handing them to a developer, and search across private company files. The model reads everything locally, so client contracts and meeting audio never leave your hardware. Real-time web research still favors cloud AI.

Here is what non-technical entrepreneurs are doing with it today:

• Document analysis and Q&A. Upload a contract, a lease, or a proposal and ask questions about it. The AI reads the document on your hardware. No copy goes to OpenAI's servers.
• Draft generation. First drafts of proposals, client emails, and marketing copy. For most writing tasks, models like Llama 4 and Mistral perform comparably to ChatGPT Plus.
• Meeting summaries. Transcribe a recording locally and generate a summary. The audio never leaves your network.
• Code review without being a coder. If you work with developers, you can use local AI to read scripts, automation, or website changes before they go to your team.
• Client research and an internal knowledge base. Ask questions across your own contracts, reports, SOPs, and past projects through a local setup, without sending any of it to a cloud service.

Now the honest limit. Cloud AI is still better for tasks that need live web search, the most current information, or the deepest reasoning, such as complex legal analysis or multi-step strategic planning. What most entrepreneurs settle into is a hybrid: local AI for sensitive work, cloud AI for general research. That is not a compromise to apologize for. It is the intelligent position, and knowing which workload belongs where is the practical skill this post is trying to give you.

Do You Need to Be Technical to Set One Up?

No. The no-terminal path now exists. LM Studio installs as a normal desktop app and gives you a working chat session in under five minutes, with no command line. Jan.ai is the closest thing to a ChatGPT replacement. Companion Core goes furthest: pre-configured hardware with a consumer interface, no Docker, no terminal, built for non-technical operators.

LM Studio is the one to start with for most non-technical users. It downloads as a standard application for macOS, Windows, and Linux. Open it and you get a searchable model browser backed by Hugging Face: filter by task, search by name, see the estimated RAM a model needs before you download it, then click Download. Per the DEV Community comparison from May 2026, you can have a working chat session in under five minutes from first launch. No command line, no config files, no server to manage.

Jan.ai is the closest local tool to a ChatGPT replacement in look and feel. Point-and-click model downloads, an intuitive chat interface, and it is fully offline-capable once a model is downloaded. It is open-source, so the privacy-minded can audit it.

Companion Core, running the Companion Hub, goes a step beyond both. The hardware ships pre-configured and ready to run. The Hub gives you one-click app installation with no Docker, no Compose files, and no terminal, and it was designed explicitly for non-technical operators. It also adds something LM Studio and Jan do not: the full self-hosted stack, AI models plus file storage, a password manager, and productivity apps, in one interface.

The honest counterpoint matters here, because overselling ease is how people get burned. Ollama, which you will see recommended widely, is command-line first with no graphical interface. It is built for developers building applications on top of local models, and it is the wrong starting point for a non-developer. And even with LM Studio, getting from zero to a working setup still takes more time and attention than signing up for ChatGPT. Faster than 2024, yes. Instant, no.

Does It Pay for Itself?

It depends on what you pay now. If you pay $20 a month for ChatGPT Plus, a budget mini-PC pays for itself but a Companion Core does not on AI cost alone. If you pay $200 a month for Pro, a Companion Core pays for itself in under two years. If a team of five shares ChatGPT, Core financing of $128 a month is roughly break-even on the subscription alone, and adds everything else.

Start with the cases where the math closes cleanly. If you are paying for ChatGPT Pro at $200 a month, that is $2,400 a year leaving your account, and a Companion Core pays for itself in under two years while eliminating the ongoing fee. If you run a team of five on ChatGPT Team, that is $25 per user per month, $125 a month, $1,500 a year. Core financing through Affirm runs about $128 a month, which is roughly break-even on the AI subscription alone, and the Core then carries everything else the Hub offers on top.

Here is the simplified version. The full three-year cost comparison, with electricity and hardware tiers, lives in our detailed cost breakdown.

Read this against your own usage rather than ours. The point of the table is that the answer changes with what you pay and how many people share it, not with a single headline number.

So here is the honest math stated plainly. If you are a light ChatGPT Plus user, a Companion Core does not pay for itself on the subscription comparison alone, and you should not pretend it does. A budget mini-PC would. The financial case for the Core is strongest when you are paying for Pro, when you have a team, or when the confidentiality of your client data has a calculable value, which is the next section.

What Do You Have to Manage If You Are Not a Developer?

Less than running a server, more than a cloud subscription. Day to day it is click-to-update models and apps from the Hub dashboard, all browser-based. Occasionally you restart the hardware or confirm backups are running. Rarely, edge cases like recovering from a drive failure or setting up outside access may need a setup guide or brief support. It is closer to a NAS than a server.

The honest breakdown, by how often it comes up and how much it asks of you:

• Low-effort, ongoing, all browser-based in the Hub. Model updates are a click of Update when a new version lands. App updates are the same click. Available storage shows on the dashboard. Adding a new app is click-to-install from the catalog.
• Occasional, and rarely technical. Sometimes you power-cycle the Companion Core, the same as restarting any device. If your home or office network changes, you adjust for it. Now and then you confirm your backups are running, which the Hub should surface.
• Rare, but worth being honest about. A few things may need help: complex integrations between apps (say, wiring a self-hosted project tool to an outside service through an API), recovering data after a hardware failure, upgrading hardware such as adding storage, or setting up access from outside your local network. That last one uses Tailscale or an equivalent and means following a setup guide, not deep technical work.

The honest conclusion: managing a Companion Core is closer to managing a NAS device than managing a server. It is not zero maintenance, and it is not a set-and-forget cloud service. It is well within reach of a non-technical business operator, as long as you go in knowing that an edge case may occasionally call for a support ticket or a short technical consult.

When Does Client Confidentiality Make It Worth It on Its Own?

When you handle data you are obligated to protect, the privacy benefit can justify the hardware by itself. Lawyers, accountants, healthcare-adjacent professionals, and anyone under NDA carry client data where a cloud leak has real cost. Running AI locally means the contract, the audio, and the client files are processed on your hardware and never transmitted. That risk reduction does not appear in the subscription comparison.

For some entrepreneurs the privacy argument is not abstract, it is structural. It matters most for:

• Legal professionals, where routing privileged material through cloud AI raises attorney-client privilege questions. We covered exactly this in how lawyers accidentally waive attorney-client privilege, and it is the sharpest version of this argument.
• Financial professionals handling client financials and tax information.
• Healthcare-adjacent work with HIPAA considerations.
• Anyone under NDA: client contracts, product roadmaps, acquisition discussions.
• Agencies and consultancies holding client source code, campaign strategy, and proprietary data.

What "running AI locally" means for these cases is concrete. A local model processes your contract document on your hardware, and no copy is transmitted to OpenAI. A local transcription model handles your meeting audio, and the recording never leaves your network. A local document Q&A system answers questions about your client files because those files are indexed locally.

This is the economics-of-ownership argument extended past the subscription line. The cost of a single data breach or a broken client trust can exceed the price of the hardware by orders of magnitude. For a professional handling sensitive client data, that risk reduction is part of the "worth it" calculation, and it is a value that never shows up in a ChatGPT price comparison.

The caveat, to keep this honest: if you do not handle sensitive client data, the privacy benefit is real but it is not urgent. It is a reason, not the reason, and you should weigh it as such.

Where Is It Not Worth It?

It is not worth it on cost alone if you use AI lightly and only pay $20 a month. Local models also trail the frontier on the hardest reasoning, have no live web access by default, and take more setup than a signup. If your work centers on deep research synthesis or complex multi-step reasoning, cloud AI may still be the better tool for those specific tasks.

These are the things that remain harder than cloud AI, and they belong in any honest answer:

• Model selection. You have to choose which local model to run. Cloud AI hides that choice from you.
• The capability ceiling. The best local models in 2026, Llama 4 and Mistral, are competitive for most tasks but do not match the frontier capability of GPT-4o or Claude Sonnet on the hardest reasoning.
• Real-time information. Local models do not have internet access by default. They answer from training data, not live search.
• Context length. Some local models have shorter context windows than cloud models, though that gap has narrowed a lot in 2026.
• Plugins and integrations. ChatGPT's plugin and integration ecosystem is broader than what is available locally.
• Setup time. Even with LM Studio, getting to a working setup takes more time than signing up for ChatGPT.

The trade-off in plain language: you are trading some capability ceiling and some convenience for complete data sovereignty and a different cost structure. For a non-developer entrepreneur whose AI use centers on drafting, document analysis, and internal Q&A, that trade-off is minimal. For someone doing deep research synthesis or complex multi-step reasoning, cloud AI may still be the better tool for those specific tasks. That is the whole answer, upside and limit, which is the only kind of answer worth trusting when you are about to spend money.

If after all of that the answer for you is yes, that you use AI enough and your data matters enough, then a Companion Core is the version built for someone who is not a developer: the hardware pre-configured, the apps one click, the terminal nowhere in sight. See the Core and what it runs and decide.

Frequently Asked Questions

Works Cited

• Running Local LLMs in 2026: Ollama, LM Studio, and Jan Compared (DEV Community)
• LM Studio vs Jan 2026 (Kunal Ganglani)
• Jan AI Complete Guide 2026 (AI Made Tools)
• Jan vs LM Studio vs Ollama 2026 (Local AI Master)
• Local AI Server for Business 2026 (Compute Market)
• Personal AI Servers 2026 (Boostely)
• Best Private Personal AI Assistants 2026 (Vellum)
• Self-Hosted AI Guide 2026 (GitHub Gist, yalexx)
• OpenAI Pricing (OpenAI)
• Companion Intelligence (ci.computer)

A budget home AI server runs about $509 over three years — hardware plus electricity. ChatGPT Plus costs $720 for the same period. At the mid-range, ownership costs roughly $1,199 vs $3,600 for ChatGPT Pro. A Companion Core at about $4,020 beats ChatGPT Pro $200 ($7,200 total) by $3,180. Hardware wins on a 36-month horizon at every tier. The question is which tier fits your use case.

Adding It Up

Most people think about AI subscriptions the way they think about a streaming service — ten, twenty dollars a month, barely noticeable. That framing is how subscriptions survive. Three years of ChatGPT Plus is $720. Three years of ChatGPT Pro at $200 a month is $7,200. The money leaves your account quietly, one month at a time.

This post runs the full cost math for a home AI server at three hardware tiers, compared to their ChatGPT equivalents. It includes electricity, names setup time honestly, and ends with a capability comparison that doesn't flatter either side.

What Does ChatGPT Cost Over Three Years?

ChatGPT Plus costs $720 over three years at $20 per month — the cumulative weight of a monthly charge most users don't notice individually. ChatGPT Pro adds significantly more: $3,600 at the $100 tier, $7,200 at the $200 tier. All plans are month-to-month with no annual discount.

OpenAI currently offers four consumer tiers:

• Free: Limited GPT-4o access, no priority queue.
• Plus ($20/month): Full GPT-4o, higher usage limits, image generation.
• Pro $100/month: 5x the Plus quota across GPT-5.4 Thinking and Deep Research. Launched April 9, 2026.
• Pro $200/month: Near-unlimited access to GPT-5.4 Pro, GPT-5.5, o1 Pro mode, and priority compute.

Three-year totals:

• Plus: $720
• Pro $100: $3,600
• Pro $200: $7,200

There is no annual plan and no loyalty discount. Every month is full price.

What Does a Home AI Server Cost at Each Price Point?

Three practical hardware tiers run local AI in 2026. Software — Ollama, LM Studio, Jan.ai, and model downloads — costs nothing at any tier.

Tier 1 — Budget (~$389) Representative hardware: Beelink EQR6 Mini PC. Capable of running 7B–13B parameter models (Llama 3.1 8B, Phi-4, Gemma 2 9B) at conversational speed. 32GB RAM minimum. Power draw around 15–25 watts under load. For most daily tasks, the practical equivalent of ChatGPT Plus.

Tier 2 — Mid-Range (~$999–$1,359) Representative hardware: GMKtec EVO-T1 (~$999) or Minisforum AI X1 Pro-470 (~$1,359). Handles 13B–32B parameter models — enough for complex reasoning, long document analysis, and serious code review. Practical equivalent to ChatGPT Pro at $100 per month.

Tier 3 — Companion Core (~$3,600) AMD Ryzen AI Max+ 395 processor. 128GB unified RAM at roughly 256 GB/s memory bandwidth. AMD Radeon 8060S GPU with 40 compute units. 2TB NVMe storage upgradeable to 16TB. 500W system running at 19 dB idle. Runs 70B to 200B-class models locally.

What distinguishes the Core from other high-end hardware: the Companion Hub marketplace (50+ local-first apps, one-click installation), Companion Memory (persistent, private, cross-session AI memory — no cloud equivalent exists), turnkey setup in roughly 2 hours without Docker, and a 1-year hardware parts warranty plus 2 years labor. Available through Affirm at approximately $128 per month.

What Is the Total Cost of Ownership Over Three Years?

Hardware plus electricity. Software costs zero at every tier.

US residential electricity averages $0.171 per kilowatt-hour. Based on realistic active-use estimates rather than 24-hour-a-day operation:

The break-even points — where hardware has cost less than the equivalent subscription — fall at month 20 (budget vs. Plus), month 10 (mid-range vs. Pro $100), and month 18 (Core vs. Pro $200).

At the budget tier, the $211 three-year savings is narrow. What you are paying for is capability at parity with ChatGPT Plus, running on hardware you own, with no subscription that can raise prices or restrict access. After break-even, the hardware keeps running. The subscription keeps charging.

What Do You Actually Get at Each Price Point?

For most daily AI use — writing, summarizing, document analysis, coding help, Q&A — a mid-range local server running a 13B–32B model is functionally indistinguishable from ChatGPT Plus. The two things a Companion Core offers that ChatGPT cannot, at any subscription tier: inference that never leaves your hardware, and Companion Memory — a persistent, private AI memory that accumulates context across sessions and apps.

What Does the Math Leave Out?

Most cost comparisons stop at the table. Four costs belong in any honest version of this calculation.

1. Setup time. A DIY mini PC build takes 10 to 20 hours to configure for the first time. A subscriber opens a browser tab. If your time is worth $50 an hour and you spend 15 hours on setup in year one, that is $750 in implicit time cost — real money that does not appear in the hardware price. Companion Core reduces this to approximately 2 hours of turnkey setup.

2. Model capability ceiling. Budget hardware runs 7B–13B models. For most knowledge work, these are capable and fast. For frontier-reasoning tasks — extended multi-step research, complex code generation, the kind of work that pushes ChatGPT Pro $200 — local budget hardware has a real ceiling. The honest answer: for 85% of daily AI use, you cannot tell the difference. For the remaining 15%, ChatGPT Pro $200 currently leads on raw model quality. And the gap is closing fast.

3. Upgrade trajectory. Hardware depreciates. In year four, the Companion Core is yours — and the hardware frontier may shift. Subscriptions include the latest model automatically, and cover the hardware that runs it. This is a real advantage of the subscription model.

4. What the Core minimizes. The Companion Core is not the cheapest path. It is the ownership path with the least friction: turnkey setup, Hub marketplace for 50+ apps, persistent memory, warranty, and a clear upgrade path. For users who want the economics of ownership without the DIY complexity, it is the option that makes the trade-off worthwhile.

For a longer look at how cloud subscription costs compound over time, see The Mirage of the Cloud: The Costs of Convenience. On building confidence in self-hosted AI without technical expertise, The Return of the Local covers the adoption question directly.

Frequently Asked Questions

Can a home AI server really do what ChatGPT does? For everyday tasks — writing, summarizing, Q&A, light coding — yes. A mid-range server running a 13B–32B model is indistinguishable from ChatGPT Plus for the majority of daily use cases. For frontier-reasoning tasks at the Pro $200 level, ChatGPT currently has an edge on raw model capability.

What if I don't want to manage a server? Companion Core is designed specifically for this. Turnkey setup in approximately 2 hours, one-click app installs via Hub, and a 2-year hardware warranty. No technical skills required.

Is the Companion Core available on a payment plan? Yes. Companion Core is available through Affirm at approximately $128 per month — subscription-comparable in the short term while building toward full ownership.

Does a home AI server work without the internet? Yes. All inference runs on your hardware. Prompts, responses, and context never leave your device. An internet connection is required only for initial model downloads and optional cloud-connected features.

What is the cheapest way to start with a home AI server? A budget mini PC at approximately $389 — the Beelink EQR6 class — running Ollama is the lowest-entry practical path. For most users, this is a capable first step into local AI.

Works Cited

• ChatGPT Plans — OpenAI
• ChatGPT Pricing 2026: All 6 Plans — TechJack Solutions
• Best Mini PCs for AI Server 2026 — PC Build Advisor
• Best Mini PC for Ollama and Local LLMs — MayhemCode
• Home Server Power Consumption — EcoFlow
• Local LLMs vs Cloud LLMs in 2026 — FreeAcademy
• Companion Intelligence Hardware — ci.computer/hardware
• Companion Intelligence Homepage — ci.computer

The GPU launched in April 2026 at $429 MSRP for the 16GB variant — the first time the sub-$500 tier came with enough video memory to run 14 billion parameter AI models at interactive speeds. 8B models had been doing it for a while. 14B models are different: noticeably better reasoning, more coherent on complex questions, and the quality step that turns "it runs" into "I'd actually use this."

Two things to establish up front. First, the "$500" in the headline refers to the GPU, not a complete computer. Adding the RTX 5060 Ti is an upgrade to a PC you already own — "under $500 to add local AI capability to an existing PC." A full system starts at approximately $1,399 at list price. If you're building from scratch, the GPU is a component, not a budget.

Second, what this post covers: what models it runs, how fast, how to set up your first local AI assistant in under 5 minutes, the honest economics, and where local AI still falls short.

What does "capable" mean for a local AI assistant on consumer hardware?

Capable means it runs fast enough that waiting for a response doesn't interrupt your thinking. At 51–71 tokens per second on 8B models and 31–34 tok/s on 14B models, the RTX 5060 Ti produces text at speeds that feel interactive — comparable to a human typist on simple tasks, and fast enough for real Q&A on 14B models.

Before the specs, here's what that looks like in practice:

You have a PDF contract, a meeting transcript, or a long article you want to interrogate. You drag it into your local chat interface. You ask: "What does section 4 commit me to?" or "Summarize the three most important points from this." The model reads it. It answers in seconds. No data leaves your machine. No usage cost. No rate limit to wait out. That interaction — document Q&A, drafting help, explaining a block of code — is what "capable" looks like in daily use.

For daily tasks, the 14B class model quality is meaningfully better than 7–8B. Responses stay coherent longer on multi-step questions, handle more context without losing the thread, and produce cleaner output on writing tasks. That quality gap is the threshold between a local AI that's a curiosity and one that's a tool.

What it's not capable of — we'll cover that in full in section six.

Why does VRAM matter for running AI locally — and how much do you need?

AI models must fit entirely in your GPU's video memory (VRAM) to run at full speed. Any part of the model that spills into system RAM runs 10–20x slower. The RTX 5060 Ti 16GB provides approximately 15.5GB usable for model loading — enough for 8B models with room to spare, and for 14B models with 5–6GB of KV cache headroom.

The 16GB threshold matters for a specific reason: Q4-quantized 14B models occupy roughly 8–9GB. That leaves 6GB or more for the KV cache — the memory used for each conversation turn. A generous KV cache means longer context windows don't force the model to compress earlier parts of the conversation.

This is what changed with the RTX 5060 Ti. The RTX 4060 Ti also shipped a 16GB variant, but at lower memory bandwidth. The 5060 Ti's 448 GB/s GDDR7 delivers approximately 50% faster LLM inference. That's the difference between 31–34 tok/s on 14B models versus under 25 tok/s on the previous generation. Both run; only one runs at interactive speed.

On price: MSRP is $429 for the 16GB variant. Street prices at launch have ranged $479–$569 due to supply pressure. Use $429 as the reference; expect to pay somewhat more near-term.

What AI models can the RTX 5060 Ti 16GB run, and how fast?

The RTX 5060 Ti 16GB runs 8B class models (Llama 3.2 8B, Mistral 7B) at 51–71 tokens per second, and 14B class models (DeepSeek-R1 14B, Qwen 2.5 14B) at 31–34 tok/s with Q4 quantization. Both speed ranges are fast enough for daily conversational use. The 14B class produces noticeably better reasoning quality than 7–8B models.

Specific benchmarks:

At 71 tok/s, text arrives faster than most people read. At 33 tok/s, responses feel like a real conversation — fast enough that you're reading one answer while formulating the next question.

The recommendation: start with a 14B model at Q4 quantization. Qwen3 14B is a solid all-rounder; DeepSeek-R1 14B has stronger chain-of-thought reasoning for coding and analytical tasks. The speed difference versus 8B is noticeable but not frustrating. The quality difference is larger, and that's what determines whether the setup actually changes how you work.

Q4 quantization compresses the model to roughly 40% smaller file size with minimal impact on most tasks. You're trading a small amount of theoretical quality ceiling for 8–9GB on disk versus 15+ GB for the full-precision version.

How do you set up a local AI assistant on an RTX 5060 Ti without coding experience?

LM Studio provides a desktop app for Windows, macOS, and Linux with a built-in model browser, one-click downloads, and a chat interface — no coding required. A complete beginner can download LM Studio, browse to a model, download it, and be chatting in under 5 minutes. Ollama is an alternative that requires one terminal command.

LM Studio — the non-developer path:

1. Go to lmstudio.ai and download for your OS. It's a standard application installer.
2. Open LM Studio. On first launch it detects your GPU and configures acceleration automatically — no driver setup required if you've already installed the NVIDIA drivers for your card.
3. Click "Search Models." Type "qwen" in the search bar. Models display with size, quantization tier, and memory requirements.
4. Download "Qwen 2.5 14B Q4_K_M." The file is approximately 8.5GB — a few minutes on most connections.
5. Click "Chat." Select your model from the dropdown. Start typing.

That's the full path. No configuration file, no package manager, no environment variables.

Model selection is where 30 minutes of reading pays dividends: understanding that Q4_K_M is the right quantization for most use cases, that context length determines how much of a document the model can hold in attention, and that the numbers in model names reference parameter count. LM Studio's model browser shows memory requirements, so you can see at a glance whether a model fits.

Ollama is worth noting as a next step: it runs a local model as an API service, meaning other tools — code editors, workflow automation, scripts — can connect to it. Setup requires one terminal command (ollama run qwen2.5:14b) but no ongoing maintenance. For a first setup, LM Studio is the better starting point.

Is buying a GPU for local AI cheaper than paying for ChatGPT in the long run?

The RTX 5060 Ti 16GB at $429 MSRP breaks even against ChatGPT Plus ($20/month) in approximately 21 months — if you cancel your subscription. For heavy local AI users who would meaningfully reduce their cloud AI spend, the economics favor ownership. For light users who run both, cloud AI's $20/month remains a low bar.

The GPU keeps running after month 21. The subscription doesn't.

The honest version: the break-even math assumes you actually cancel the subscription. Most people run local AI alongside cloud AI rather than replacing it entirely. The economics work best for someone already spending $20–$50/month on AI subscriptions who is willing to route the private, repetitive, and high-volume tasks to local — not add a GPU on top of an unchanged subscription stack.

The value that doesn't show up in the break-even calculation: no per-query billing, no rate limits, works offline, data never transmitted. For tasks involving sensitive documents or intermittent connectivity, those properties have real worth independent of the ChatGPT cost comparison.

Hardware you own costs nothing to run after purchase. The value compounds the longer you use it. The subscription compounds in cost.

See also: Stewardship Systems: Building Ethical & Sustainable AI Operations

What can't local AI models on the RTX 5060 Ti do that cloud AI can?

A 14B model running locally is not a match for GPT-4o or Claude Sonnet on complex multi-step reasoning, advanced coding, or tasks requiring current world knowledge. Local AI wins on privacy (data never leaves your machine), offline availability, and zero per-query cost. It loses on the reasoning tasks where frontier model scale matters.

Specific failure modes:

Complex multi-step reasoning: Tasks requiring many interdependent constraints simultaneously — intricate logical chains, complex legal analysis, planning with many variables — show where parameter scale matters. A 14B model gets partway and loses coherence where a frontier model doesn't.

Advanced coding: Local models handle function writing, code explanation, and isolated bug fixes well. They struggle on large-codebase debugging (10,000+ lines) and architectural decisions where filling the context window early is a hard constraint.

Current world knowledge: Local models have a training cutoff. They don't know what happened last week. Cloud AI with browsing can look things up in real time; a local model running offline cannot.

The right frame isn't "worse." It's different tool for different tasks. Local AI is the right call for private tasks, offline work, and high-volume repetitive operations where per-query costs accumulate. Cloud AI is the right call for tasks requiring frontier reasoning quality or live world knowledge. The RTX 5060 Ti makes the first category accessible in a way it wasn't at this price before.

Is there an easier way to manage local AI apps on a home server?

Companion Hub installs and manages self-hosted AI applications on your own hardware with a one-click interface — no command-line configuration, no manual container management. For a machine built around the RTX 5060 Ti, Hub handles the software layer above the GPU so you're running apps rather than maintaining an infrastructure stack.

The RTX 5060 Ti handles the compute. What sits above it — managing models, updating applications, running multiple self-hosted tools, keeping everything configured — is a separate layer of friction. The hardware threshold has been crossed; the software layer is the next one.

Hub is designed to run on hardware exactly like this: a machine with a capable GPU and always-on home network connectivity. The apps in Hub's marketplace install into that hardware environment without requiring you to understand what's running underneath — the same way apps install on a phone.

The GPU is what makes the hardware viable. Hub is what you do with viable hardware once the configuration overhead isn't the bottleneck.

Download Hub at hub.companionintelligence.com

See also: The Return of the Local: Building Confidence in Self-Hosted AI

The threshold

Local AI has always been expensive or complicated — usually both. The RTX 5060 Ti 16GB is the first time the price came down to a single GPU upgrade that a person with an existing PC can make, while the software simultaneously reached the point where 5 minutes from install to first response is real.

The question the headline asked is answered: yes. You can now run a capable home AI assistant for under $500 — where "capable" means 14B model quality at speeds that feel like a conversation, on the tasks you'd actually use it for daily.

The ceiling is real: a local 14B model is not GPT-4o. Know the tasks you want it for, run both for a month if you're on the fence about the economics, and find your own answer. The hardware is now affordable enough that the experiment has become low-stakes.

Frequently Asked Questions

Do I need a whole new PC or just the GPU?

Just the GPU, if you already have a compatible PC. The RTX 5060 Ti requires a free PCIe x16 slot and a power supply rated 600W or higher — check your case documentation before purchasing. A complete new system starts at approximately $1,399 at list price.

Is LM Studio free to use?

Yes. LM Studio is free for personal use. The app, model downloads, and local inference are all zero cost — no subscription, no API keys, no usage limits.

Can I use the RTX 5060 Ti for both gaming and local AI?

Yes, but not simultaneously. Both gaming and local AI inference use GPU resources; run one or the other. Switching between them takes seconds, but VRAM contention makes running both at once unreliable. The GPU handles each task independently when the other isn't running.

What 14B model should I start with?

Qwen 2.5 14B or Qwen3 14B are solid starting points for general Q&A, drafting, and document work. DeepSeek-R1 14B is worth trying if coding help and chain-of-thought reasoning are your primary use cases. Both run well on the RTX 5060 Ti 16GB at Q4_K_M quantization. Start with Qwen 2.5 14B as an all-rounder.

Does the RTX 5060 Ti also support image generation (Stable Diffusion)?

Yes. 16GB VRAM provides comfortable headroom for Stable Diffusion XL and similar image generation models. LM Studio is text-only — for image generation on the same hardware, use ComfyUI or Automatic1111. Both are separate installs that can run on the same card.

Last updated: [YYYY-MM-DD]

Works Cited

• S1: NVIDIA, "GeForce RTX 5060 Ti Product Overview," 2026. [NVIDIA.com] (C1 — launch date, MSRP, specs)
• S2: Tom's Hardware, "RTX 5060 Ti Review: LLM and Gaming Performance," 2026. (C1, C11 — VRAM, inference benchmark comparison)
• S3: modelfit.io, "RTX 5060 Ti LLM Inference Benchmarks," 2026. (C3 — tok/s, model list)
• S4: localscore.ai, "Consumer GPU Local AI Benchmark Database," 2026. (C3 — speed ranges, Q4 quantization comparison)
• S5: LM Studio, "LM Studio Documentation: Getting Started," 2026. (C6 — setup path)
• S6: Ollama, "Ollama README," GitHub, 2026. (C7 — terminal command, API access)
• S7: OpenAI, "ChatGPT Plus Pricing," 2026. [openai.com/chatgpt] (C8 — $20/month)
• S8: DeepSeek, "DeepSeek-R1 Model Card," 2026. (C4 — model specs)
• S9: Qwen Team, "Qwen3 Technical Report," 2026. (C4 — model specs)
• S10: PC Guide, "RTX 5060 Ti 16GB Street Price Tracker," 2026. (C5 — retail range, full system cost)