Personal AI Should Actually Be Personal

Last week, Amazon blocked Meta’s Muse from its retail site. People trying to shop got a popup telling them that continued access by an unauthorized AI agent violates Amazon’s conditions of use.

Nobody asked the person whose agent stopped working. I’m a customer of both companies, but Amazon blocked Meta, and my access disappeared as a side effect of a dispute I had no say in. And shopping is one of the main things these agents are for.

We’ll see more and more of these platform wars as companies battle to own the consumer and we’ll be caught in the crossfire.

An agent hosted by a platform has two principals. It works for you, and it works for the company that built it, and we know which one takes priority. What we’re all calling personal AI isn’t personal at all.

None of this is a knock on Muse or Instinct. They’re beautiful, thoughtful products, and there’s going to be a whole range of choices here, which is healthy. I just think one of those choices has to be one you actually own: a productized version of an open, independent, opinionated, personal AI service, built on OpenClaw.

What personal should mean

Here’s what I think personal AI has to mean. The rest of this is the case for each.

1. You own it. Open source, your information in storage you control, and you can take everything with you.

2. One principal. Nobody else influences what it does or recommends. The only money that changes hands is yours, for the product.

3. You know the deal. The business model and what happens to your data are plain and don’t change without your consent, and when it chooses for you, you can inspect what it weighed and why.

4. You choose where it runs. Your laptop, your Mac Studio, or a machine you lease, and you can move.

5. It’s delightful. Beautiful, simple and easy, because otherwise people will go back to the easier product.

It isn’t your agent

Muse users are opted in by default to having their conversations used for model training. Instinct’s terms grant a perpetual, irrevocable license to your materials, including for training, covering screen captures and keyboard input. None of this requires anyone to misbehave. They tell you what they do with your information, and what your agent can do on any given day depends on their relationships with other companies. It’s theirs, and you get to use it in exchange for creating value for them.

That matters more here than anywhere, because personal AI is maximalist on information. Every additional thing it knows about your calendar, your inbox, your finances, your relationships makes it better, and that’s exactly the information you should least want to hand to a third party.

If you don’t pay for the product, you are the product. Brad (@altcap) made the case last week (opens in new tab) that a personal super assistant in everyone’s pocket is inevitable, and I agree. His answer on trust was that paid versions won’t steer you, free ones will take fees, and companies know trust matters so they’ll solve for it. That’s the part I see differently. So far these platforms, while maybe well-intentioned, haven’t proven themselves to be trustworthy stewards of our experience and our data.

Steering matters more here than it ever did with search. On Google you get ten blue links, and you make an active choice: the paid result at the top because it’s well targeted, or one of the organic ones. With an agent, you mostly just want the outcome. You’re delegating the choice, so any nudging happens silently, through an algorithm you don’t understand. That’s why the agent has to work on your behalf with your preferences baked in, and let you inspect how it decided whenever you want to.

We knew there was a gap between what these assistants cost to run and what we pay for them. It’s a big one. Anish (@illscience) is hearing that personal assistants cost $3k to $7k per user per year, which matches what I’m seeing in my own OpenClaw usage. At $20 a month, someone is covering the rest. Meta says Muse conversations don’t go to its ad systems, for now. But with a gap that size and an ad business at the core, how would one expect Meta to get paid back on that subsidy long term?

Why this matters

I’ve seen the power in my own use of AI tools and OpenClaw. Once it’s personalized with your information and your tools, one person gets an enormous amount of leverage. Personal AI for everybody is going to be hugely empowering.

Every major platform sees how critical this is, and I think it matters even more than search and social. Personal AI is the path to all of your personal information and all of the commerce that flows through you, which is why everyone with real distribution is going after it. That’s exactly why we need an independent, open, personally owned alternative that’s just as capable and easy for anyone to use, where you understand exactly what the business model is and what happens to your data.

Personal ownership is what makes it more powerful

In the history of digital products, usage flows to whatever is simplest and provides the best user experience, regardless of privacy considerations. People want low friction and high quality, and privacy risk rarely makes the list. So anything owned and private is fighting uphill. Normally the company holding the most data on you also has the better product. Facebook knows more about you than you do, and that’s why it works so well.

Here it reverses. Because you own it, you’ll give it everything: every account, every document, every corner of your life you’d never connect to Meta. And that information compounds, which is why it’s worth investing in something you own sooner rather than later. Whichever product has the fullest access has potential to be the better product, and ownership is what makes fullest access possible. This is one of the very few cases where the private version can deliver the better product, as long as the rest of the experience is as good as what the commercial products ship.

OpenClaw feels like Linux. We need Apple experience.

OpenClaw is an amazing project, I use it all day, and it’s the inspiration for a lot of what’s shipping now, Muse included. But to the average user it still feels like Linux. It takes too much setup out of the box, and to run it well you need to be comfortable in a terminal, understand context and caching, and know how to configure models. That’s fine for people like me. It’s a non-starter for everyone else.

What we need feels like Apple. You choose a few options, you put in a credit card, and it works. You own the whole stack, it’s beautifully done, and it isn’t controlled by one entity whose incentives might not line up with yours.

The personal AI stack

Here’s the stack I think it takes, paid for by the user instead of by their information, and where each piece is today. Your information is only readable inside the computer you control, and anything that leaves it is deliberate, minimal and visible to you. You and only you can steer it and control it.

Clients. Beautiful, modern apps on your phone and desktop, plus the messaging apps you already use. This is where Muse and Instinct are ahead and leading the way. OpenClaw has its own native, web and mobile apps, but they’re built for power users, with lots of buttons, features and settings, rather than a focused, Muse-like experience.

Harness + Connectors. The part that acts: tools, connections to your accounts, your identity and credentials, payment options, what the agent is allowed to do, approvals for anything that matters, and a log of everything it did. OpenClaw and Hermes are pioneering here today and inspired Muse’s architecture. This is the core of the orchestration and needs to be owned and inspectable by the user.

Memory. What comes in through your connections (email, calendar and the rest), plus your preferences and what it learns about you over time, in a form you can read, correct and take with you. @karpathy’s LLM wiki (opens in new tab) is the clearest version of this so far: the model compiles your information into plain markdown files it keeps up to date, and you can read and edit every page. @garrytan’s GBrain (opens in new tab) is a step in that direction: the same markdown-first idea, built as memory for agents like OpenClaw and Hermes. This is where the value compounds over time, which is exactly why it has to be yours.

Models. I imagine two tiers. A default open-weights model tuned for personal agentic workloads handles almost everything, cheaply, inside your boundary. A frontier model gets called deliberately for frontier-level judgement (which will likely always be ahead), with only the context it needs. Most of the time the question is only whether anything needs doing, and a small model can answer that far more cheaply: one breakdown (opens in new tab) puts it at about $2 a month per user today, versus about $46 on common defaults. Capable small open models exist today; the routing between the tiers is mostly hand-built.

Your computer. Where the harness, memory and default model live: your own machine, or a private one leased from someone whose only business is leasing it and who can’t read what’s on it. Both exist today, but the private hosted option is still built for developers. A box in this class rents for about $10 a month (opens in new tab). It should be as easy as Muse giving each user their own VM, except the VM is yours.

What I would love to see built

Every layer has someone working on it. Nobody has put them together into something that works and feels like Muse, but where you give it a credit card instead of all of your personal data.

That’s the product I want: a productized version of OpenClaw, with the same open-source energy and one principle, that the people using it own it. Where OpenClaw deliberately takes no point of view on models, infrastructure or hosting, this version bakes those choices in, keeps every one of them reversible, and puts focused, Muse-like surfaces on top: a mobile app, a desktop app, the messaging apps people already live in.

Owning it doesn’t mean nobody gets paid; someone has to make it beautiful and keep it evolving. It means the only money anyone makes comes from the user, and the user can leave with everything: pay for updates, keep what you bought, and buy the software (the old school way!), the machine and the model calls from people whose only business is selling you those things.

OpenClaw already has the foundation, chaired by @davemorin. What we need is the company. If you’re building it, or know who is, I’d love to hear from you.