Why I build my own AI OS

My wife Kim had two invoices from a dentist visit that needed claiming on my Bupa cover. She emailed them to me and asked if I could sort it. I forwarded the job to Otto, my AI OS, in one line:

It found the email, opened the Bupa site in its own browser, logged in with its own password manager and filled in the claim. The only thing it needed from me was the code Bupa texted to my phone. About five minutes later Kim got a confirmation and messaged me to say thanks. I told her not to thank me, thank Otto.

Otto isn’t a product I bought. It runs on its own Mac mini, and I built it.

There are plenty of good agent tools around now: Claude Code, Hermes, OpenClaw and what feels like a new one most weeks. I use some of them. But I’ve always built my own harness, and I’m more convinced than ever that it’s the right call.

Every agent tool is the same four parts

Strip any of these tools back and you find the same four things. I call them CASA: context (what the AI knows about you and your work), access (the tools, accounts and systems it can reach), skills (the ways of doing things you’ve taught it) and automations (what runs without you asking).

The industry is quietly agreeing. In December 2025, Anthropic handed MCP, the standard way for an agent to reach other tools, to a new Agentic AI Foundation under the Linux Foundation, co-founded with OpenAI and Block. OpenAI’s AGENTS.md, a plain file telling an agent how to work in a project, went in alongside it. The same month Anthropic opened up Agent Skills as a format any tool can use. Access, context and skills are each turning into something every harness can read.

When the parts are standard, the tool around them matters less. What matters is what you put into each part, and who controls it.

Built for everybody, built for no one

Software has always been built to fit as many people as possible, because building it was expensive and the cost had to be spread across thousands of customers. That made sense. It also meant every tool was a compromise, designed around an average user who doesn’t exist.

That cost has collapsed. Andreessen Horowitz worked out that the cost of an AI model at a given level of performance falls about tenfold every year. Epoch AI found the price of reaching the same benchmark score falling somewhere between 9 and 900 times a year, depending on the test. And those models are now doing most of the building. When building is that cheap, adapting yourself to a product designed for someone else stops making sense.

I run the same underlying system for my personal life and for EmbedAI, with small differences where the two need to differ. Both are built around the way I work. If I want something to behave differently, I change it, usually by asking for the change.

It compounds

A few weeks ago I was playing padel at a new club in Wimbledon, and the managers asked me about marketing. So I built them a local SEO rank tracker. Every week it checks where the club shows up on Google Maps across the surrounding neighbourhoods, how it compares with the other padel clubs nearby, and how everyone’s reviews are moving. The first check had the club top of the map for “padel” in four neighbourhoods and missing entirely from three more, which told them exactly where to start.

It didn’t take long, because my AI OS already had most of what it needed: the context on how I like things built, access to the SEO data and the hosting, and the skills for doing SEO work properly. And now that it exists, the system knows how to do local SEO for a small business, and I own a tool I can set up for anyone.

That’s the part an off-the-shelf tool can’t give you. Every build leaves the whole setup a bit better, and the next one gets easier. Otto now has more than a hundred skills, and the ones I use most exist because some earlier job needed them.

Compounding also means timing matters. The earlier you start, the more you’ve built up by the time everyone else gets going.

The models keep getting better underneath it

Mine is model agnostic. When a better model comes out, I plug it in and everything carries on working the same way, only better.

The improvement is fast. METR measures how long a task an AI agent can complete on its own, and found that length doubling about every seven months since 2019. On the data since 2024, it’s closer to every three months.

That’s why owning the other parts matters. The model is the bit that improves on its own. The context, the access and the skills are the bits only you can build, and they’re the bits that make a better model useful to you specifically.

It’s also why I only recently started trusting it with things like the Bupa claim. For a long time I wasn’t confident an agent could operate a browser properly. That’s changed, and I didn’t have to rebuild anything to get the benefit.

The fair objection

Building your own isn’t free. Otto’s Mac mini is £899 new for the base model, before any model subscriptions. It takes longer to get going, and when something breaks, it’s yours to fix. Otto has its own machine, its own browser and its own 1Password vault with only what it needs: not my banking details. Working out those limits is my job, not a vendor’s.

And the ready-made tools are good. If you just want an agent that can do a few things for you, use one. For most people that’s the right first step.

But I’d still treat it as a first step. Start with Claude Code and a folder of context about you, and take control of each part as it starts to matter: write down how you like things done, connect the accounts it needs, turn the jobs you repeat into skills, and let the boring ones run on a schedule.

The part that lasts

The harnesses will keep changing. Some of the ones people use today won’t exist in two years, and the models underneath will be replaced several times over.

What you’ve put into it stays: what it knows about you, what it can reach, and how you’ve taught it to work.

That’s the part worth owning.

Sources: Anthropic, Donating MCP to the Agentic AI Foundation (Dec 2025) · SiliconANGLE, Agent Skills as an open standard (Dec 2025) · a16z, LLMflation (Nov 2024) · Epoch AI, LLM inference price trends (Mar 2025) · METR, Time Horizon 1.1 (Jan 2026) · Apple UK, Mac mini (checked 19 Oct 2026).