Computers for AI agents
Each machine is a small Linux desktop with a browser that keeps its logins, a vault the model can’t read, and an MCP endpoint. Stop it tonight and it starts tomorrow exactly where it left off. One for your agent, or one for every customer.
Free to sign up · Add a card to run machines and get $1 of credit · Billed by the minute
research-bot
Running
$ sudo apt install -y gimp
Setting up gimp (2.10.38) …
$
The same machines power a single agent or every customer of your product.
For agent builders
claude mcp add --transport http research-bot \ https://burrowbox.dev/api/machines/a1b2c3/mcp \ --header "Authorization: Bearer tmm_…"Start free Open dashboard
For SaaS platforms
# a machine for your customer "acme"
curl -X POST https://burrowbox.dev/api/machines \
-H "Authorization: Bearer $BURROWBOX_KEY" \
-d '{"name":"acme","size":"tiny","ttlMinutes":60}'
Get an API key
Create an API key
Each machine is a full, private Linux computer, small enough to run dozens at once.
A quiet little Linux desktop. Agents see it through screenshots and accessibility trees, then click, type and scroll through the same MCP tools.
Its own persistent browser profile. Cookies, logins and tabs survive restarts, so your agent doesn’t start from zero every morning.
Logins and 2FA secrets live in an encrypted vault. The agent asks the machine to log in, and the password never reaches the model.
apt packages, .deb files, AppImages, Flatpaks or a shell script. Whatever the agent installs is still there next time.
Stream the whole desktop, one app window, or just the browser. Jump in with your own mouse and keyboard whenever you want.
Always on, 30 minutes, or gone when the job is done. You only pay while it runs.
Stop a machine at 6pm and start it at 9am. Same windows, same tabs, still signed in.
Prepaid credit, no subscription. New accounts start with free credit.
Stopped machines keep their disk for a small hourly fee. When your credit runs out, machines stop and keep their state.
A small, isolated Linux computer running in a container: a lightweight desktop, a browser, a credential vault, and a daemon that exposes everything over MCP. It has its own disk that persists until you destroy the machine.
Yes, that’s what platforms use Burrowbox for. Create machines through the API with your key (one per customer, or per task), give each customer’s agent the token scoped to its machine, and set a schedule so idle machines turn off and keep their state. Usage is metered per machine.
Anything that speaks the Model Context Protocol over HTTP, including Claude Code, Claude Desktop, and custom agents built on an MCP client. Each machine has its own endpoint and a token scoped to that machine only.
They’re encrypted on the machine with a key held by the platform, not on the machine’s disk. Agents fill login forms with browser_login or vault_type_secret, which inject the secret without ever returning it to the model.
Its entire filesystem is snapshotted (files, installed apps, the vault and browser cookies), along with the list of open windows and tabs. On the next start the snapshot is restored and your windows and tabs reopen where they were. While a machine runs, a save point is also taken every 30 minutes.
Running machines are stopped, and their state is kept. Top up and start them again. Nothing is deleted unless you choose to destroy a machine.
Yes. Open a machine to stream the full desktop, a single app window, or just the browser in real time, and use your own keyboard and mouse at any point.
For your own agent, or for every customer you serve.