Setting up an agent that’s actually yours

Most AI agents you can sign up for today are a chat window with a timer bolted on. They live in someone else’s shared runtime, they come alive when you type, and if you ask what they did overnight or where your data sits, the answer is a shrug. We launched bootkey because we wanted the opposite: an agent you can boot in about two minutes, that lives on a computer of its own, thinks with a model you choose and pay for directly, and that you can log into like any other machine you run.

This is a walk through what setting one up looks like, and why each step is there.

Six choices that shape how it behaves

Booting an agent takes six decisions. None of them are decoration: each one changes what the agent notices, how well it thinks, what it remembers and what it costs. You can change nearly all of them later.

Temper: how alive should it be? A tool wakes up when you talk to it or when one of its duties is due. A worker checks in on a steady beat, every ten minutes to an hour, does what’s due and goes back to sleep. An alive agent decides for itself when to wake next. Every time it wakes it ends with a line like “next wake: 40m, waiting for the deploy”, and it sleeps until then: sooner while something is in progress, hours later when all is quiet. Of everything we built, this is the closest we’ve come to an agent that manages its own attention.

Job: what is it for? Pick a template or start blank. There’s an immune system that watches your servers, watchers for sites, pages, releases and backups, a research assistant, a morning briefing, a repo keeper, a tutor and more. The watchers check first and think second: a small script looks for a change, and the model only wakes if it found one. So you can watch a site every ten minutes and pay nothing until the day it breaks.

Mind: which model thinks for it? The picker shows where each model ranks on LMArena’s public agent leaderboards: completing agent tasks, not inventing tool calls, getting shell commands right the first time. An agent earns its keep by acting, and a model that writes lovely prose can still be a clumsy operator. You can also give the agent a cheaper model for its routine wake-ups and keep the strong one for real work, which is often the biggest saving.

Brainstem: your key. The mind runs on your own OpenRouter or Venice account. Your prices, your limits, your spending cap. We don’t resell tokens and we don’t add a margin on them.

Memory: how it remembers. Plain files you can read and fix, an Obsidian vault you can sync to your laptop, or a semantic store that recalls the right facts by itself. Agents with a rhythm also dream: in the small hours they go back over the day, merge what’s duplicated, link what’s related and let go of what’s stale.

Body: where it lives. Frankfurt, New York, San Francisco or Sydney. Each agent gets its own box with one CPU, a gigabyte of memory and 25 GB of disk. The region also sets its clock, so a 7:00 briefing arrives at 7:00 where the agent lives.

Two minutes later, it’s awake

Press boot and the console shows the agent coming up step by step. A couple of minutes later it’s running, and its Mind panel starts telling you what it’s doing: thinking, awake, in light or deep sleep, or dreaming. Its duties are a list you can read and edit, and every scheduled job says who set it: you, the template, or the agent itself.

You get the keys. All of them.

This is the part we care about most. You can log into your agent’s box over SSH, as the agent’s own user and as root. Its memory is files you can open. Nobody at bootkey keeps a key on the box unless you ask for support access.

If you want it to look after your other servers, it gets a login of its own on each one, without sudo and never with your password. The first time it connects it memorizes that server’s fingerprint, and it refuses anything that later pretends to be the same machine. API tokens it needs go into a vault: encrypted, never shown to anyone again, and handed to the agent’s tools as environment variables.

A lot of hosted AI treats people like children: a locked box, a cheerful avatar, no way to see what it actually did. We’d rather show you everything and let you decide.

Or let your own AI do the setup

On the boot page there’s a button called Copy agent prompt. It copies a short prompt with a one-hour setup key inside. Paste it into Claude Code, Codex, Cursor or a local model, and that AI reads our guide, talks the options through with you, then boots and configures the agent itself. It can set up its duties, its notes, even its access to your servers.

The key is deliberately narrow. It boots one agent and sets that one up. It can’t touch billing, your team or your model keys, and it can’t delete anything. Every change it makes appears in the agent’s activity log, signed as your AI assistant. Watching one AI interview you and then build another is more fun than it has any right to be.

What it costs, out loud

There are two meters, and they stay separate on purpose. The body is $9 a month per agent, charged by the hour while the box exists, or 50 cents a month while it hibernates with its disk saved. The mind is tokens on your own key at your provider’s prices. Each agent’s metabolism panel shows both, ticking in real time, so you can see what a busy afternoon costs compared with a quiet night.

And when it’s time to go

Agents can hibernate and wake again a few minutes later with everything they knew. If you destroy one, you’re first offered its soul and memory as a download: its instructions, its notes, its duties, everything it learned. Then it gets a proper send-off. That’s a small thing, and we didn’t want to leave it out.

What’s next

Agents spend a surprising share of their time writing code: a script to check something, a fix to a config, a small tool for a job. So we’re giving every agent better hands for code. We tested four coding harnesses on the same boxes our agents live on, and published what we found in Same model, different hands.

If you’d like to try it, boot an agent. It takes about two minutes.