"Private AI agent" means four completely different things depending on who is selling it to you. To a defence contractor it means a server in a locked room. To someone on their phone it means the chat app is not reading their messages. Both are legitimate, and the gap between them is where most buyers get sold the wrong thing. This guide separates the levels, says plainly what each one costs, and names which one you actually need.

Four tiers of AI agent privacy, from a model running locally on a laptop through to a consumer chat app on shared servers
Where your data goes changes at every level, and so does who is allowed to read it.

The four levels of "private", in plain terms

The word does almost no work on its own. What matters is two separate questions that vendors deliberately blur together: where does my data physically go, and what is anyone allowed to do with it once it is there. A local model answers the first question perfectly and the second one trivially. A commercial API can answer the second question strongly while the first answer is still "someone else's servers". Neither is a lie. They are different guarantees.

LevelWhere your data goesWho can see itCostSetup effortBest for
1. Local / on-device Never leaves your machine Only you $0 software, your existing hardware Medium, a command line install Anyone handling material that legally cannot leave a device
2. Stateless private cloud Vendor servers, deleted after the request Nobody, per the vendor's published design Bundled into the device or OS None, it is already on Everyday personal use on a phone or laptop
3. Commercial API or platform Vendor servers, retained per their terms The vendor, under contract, not the training pipeline From $0 to roughly $20 to $30 a month Low Real work with normal business confidentiality
4. Consumer chat plan Vendor servers The vendor, and possibly the training pipeline Free, or a personal subscription None Low-stakes drafting and questions

Read that table twice before you shop. Most people searching for a private AI agent picture level one and end up buying level four, because level four is what the app stores put in front of them.

Level 1: running the model on your own computer

This is the only level where the privacy claim needs no trust at all. The model file sits on your disk, the inference happens on your processor, and no network request leaves the machine. Ollama is the most common way in: it runs open models on your own computer, it is free to start, and it exposes a local API that agent frameworks and editors can call, so you can point existing tooling at it instead of a hosted provider.

We keep a local setup running alongside our hosted work, mostly to keep ourselves honest about the tradeoff. Two things are consistently true. The first is that the privacy really is absolute in a way no policy document can match, and for a certain class of document that settles the argument on its own. The second is that the ceiling is lower and the friction is real: you are choosing the model, you are the one who notices when it needs updating, and the quality gap against a frontier model is obvious on harder reasoning. For summarising your own notes it is more than enough. For an agent that has to make a judgment call you would otherwise make yourself, you will feel the difference.

The honest cost is not the software, which is free. It is that you become the operator. If that sentence sounds like a hobby rather than a chore, level one is genuinely the best answer available to you and you should stop reading roundups and go install it. If it sounds like a second job, keep reading.

Level 2: private cloud that forgets you

The middle option is newer and it is the one most people underrate. Apple's Private Cloud Compute is the reference design: requests go to Apple servers when a task is too large for the device, but the published guarantees are that the data is used only to fulfil that request, that it is deleted once the response is returned with nothing retained in any form, and that it is never available to Apple, including staff with administrative access to the production hardware. Apple also had the system independently inspected, which is the part that separates a design from a promise.

What you give up is control over the roadmap rather than control over the data. You cannot point it at an arbitrary model, and you cannot make it do a task the vendor has not built. What you gain is that the hardest privacy engineering has been done by someone else and you did not have to become a system administrator to benefit from it. For personal use on hardware you already own, this is usually the highest privacy-per-unit-of-effort available.

Level 3: commercial APIs and platforms that do not train on you

This is where most real work happens, and where the marketing gets slipperiest, because "we don't train on your data" gets presented as if it were the same as "your data never leaves your building". It is not. It is a narrower and still valuable promise.

The commitments here are worth reading precisely. Anthropic's privacy documentation states that by default it will not use inputs or outputs from its commercial products, naming Claude for Work and the Anthropic API, to train its models, and that this changes only if you explicitly submit feedback or otherwise opt in. Note the word commercial. Anthropic routes its consumer plans to a separate policy document, and that split is the single most important detail in this entire guide: the same company can have a strict policy on its business product and a looser one on the free app, and most people are using the free app. Before you trust a claim you read in a comparison table, check whether it was written about the tier you are actually on.

That distinction is also the reason we tell people not to shop on the word "private" at all. Shop on the specific sentence in the specific terms that apply to your specific plan. Everything else is decoration.

What a private AI agent actually costs

Answers to this question tend to come from enterprise vendors, which is why they are usually wrong for individuals. You will read that private deployment breaks even somewhere north of hundreds of thousands of monthly transactions. That is true and it is irrelevant if you are one person with a laptop.

Here is the individual and small-team version. Running locally costs nothing in software and nothing per month; the cost is hardware you probably already own plus your attention. Stateless private cloud on a device you already have costs nothing extra, because it is bundled into the operating system. Commercial platforms and agent tools generally start free and then sit in the $20 to $30 a month range for a single user, which is the same band the wider agent category has settled into. Enterprise on-premises deployment is the one that carries a large capital cost and a dedicated operator, and almost nobody reading this needs it.

On Gravity the free tier covers one agent at $0 a month, and paid plans start at $20 a month with $20 of usage included, with the option to buy extra usage beyond your plan. We are a hosted platform, so we sit at level three: your work is not training data, and it also does not run on your hardware. If your requirement is that nothing may leave your machine, we are not the right tool and a local setup is. We would rather say that here than have you find out after paying us. For a wider price comparison across the category, our roundup of the cheapest AI agent platforms works through the total cost of each route, and how AI agent pricing works covers the billing models themselves.

Building your own private agent

A large share of the people searching this topic are not looking to buy anything. They want to build the thing themselves, and the search history is full of people describing it as wanting their own Jarvis.

The realistic path is shorter than it looks. Run a local model through Ollama, pick an agent framework to handle the loop of planning and calling tools, and give it access only to the specific things it needs. Our guide to open-source AI agent frameworks compares the main options for that middle layer, and if you are wiring an agent into something you already run, connecting an agent to a private API covers the mechanics without exposing more than you meant to.

The part worth saying out loud is that the model is the easy half. The work that actually consumes your evenings is everything around it: deciding what the agent is allowed to touch, handling the runs that fail halfway, and noticing when it has quietly been doing the wrong thing for a week. That is the same work a platform is doing on your behalf, which is the entire argument for paying for one, and the entire argument against paying for one if you enjoy doing it yourself.

How to choose, in four questions

Work down this list and stop at the first yes.

Is there a legal or contractual rule that this data cannot leave a device? Then level one, local, and the decision is already made for you. Is this ordinary personal material you would simply rather nobody read? Level two covers it with no setup and no monthly cost. Is this business work with normal confidentiality expectations? Level three, and your job is to read the terms for your actual plan rather than the marketing page. Is this low-stakes drafting? Level four is fine, and pretending otherwise wastes money.

The common mistake runs in one direction. People overbuy privacy on material that did not need it, spend a weekend on a local setup, find the quality frustrating, and go back to the consumer app for everything, including the sensitive document that started the whole search. A correctly chosen level three beats an abandoned level one every time. If you are weighing specific tools for personal use, our comparison of the best personal AI agents covers which ones have a genuinely private posture, and the free tier comparison shows what you can test without paying. For the security practices that apply once an agent is actually running, AI agent security best practices covers permissions and blast containment, and if the underlying concept is still fuzzy, start with what an AI agent is.

Frequently asked questions

Is there an AI that is completely private?

Yes, if you run it on your own computer. An open model served locally through a tool like Ollama never sends your input anywhere, so there is no policy to trust and no vendor to audit. Every other option, including the strongest private cloud designs, asks you to trust a published guarantee instead of a physical fact. That trust is often reasonable, but it is a different kind of assurance.

How much does a private AI cost?

For an individual, usually nothing or about $20 to $30 a month. Local models are free to run on hardware you already own. On-device and stateless cloud features are bundled into the operating system. Hosted platforms typically start with a free tier and move into the $20 to $30 range for one user. The six-figure numbers you see quoted belong to enterprise on-premises deployments and do not apply to personal use.

Are there AI agents for personal use?

Yes, and the category has grown quickly. The practical split is between assistants bundled into software you already pay for, standalone personal agent tools, and platforms where you describe a recurring task and get the finished result back. Which is right depends less on features than on whether you want to build the automation yourself or just want the outcome.

Is there a free personal AI assistant?

Several. Local models through Ollama are free. The assistant features built into major operating systems and existing subscriptions cost nothing extra. Most hosted platforms, Gravity included, have a free tier: ours covers one agent at $0 a month, which is enough to test whether the thing works for your task before any money is involved.

Does "we don't train on your data" mean my data is private?

It means one specific thing: your inputs are not fed into the vendor's model training. It does not mean the data never reached their servers, and it does not always apply to the plan you are on. Providers frequently run one policy for commercial products and a different one for free and personal tiers. Check which document covers your plan before treating the claim as a privacy guarantee.

Can I run a private AI agent without technical skills?

Not at level one. Local models require a command line install and ongoing upkeep, which is a real barrier and an honest one. Levels two and three are designed for people who are not going to administer anything, which is why we usually point non-technical users there rather than talking them into a home server they will abandon in a fortnight.

Sources