Agent computers
Some AI apps now run their own computer for you. One command sets up your dev environment on that computer: your tools, instructions and skills, your history and the session board. It needs a shell, so which products it can reach depends on whether they have one.
Last updated
Which products#
Nothing here is labeled "works" until it has run on the real product. Each row is dated.
| Product | Shell | Level | Notes | Checked |
|---|---|---|---|---|
| Meta Muse | Debian | Untested | Outbound requests need your approval in Sentinel: allow 0bridge.dev, and the AI provider your agent uses, when asked | 2026-10-01 |
| Manus | Ubuntu | Untested | Ubuntu shell | 2026-10-01 |
| OpenAI Dots | Unknown | Unverified | We don't know yet whether Dots runs shell commands | 2026-10-01 |
| Grok Bot | Unknown | Unverified | We don't know yet whether Grok Bot runs shell commands | 2026-10-01 |
| Instinct | None | Not supported | No terminal or MCP yet: Instinct | 2026-10-01 |
| Another agent's computer | Linux or macOS | Untested | Any computer where your agent runs shell commands | 2026-10-01 |
Untested means the product has a shell and we believe the command runs there, but we haven't run it on the real product. Unverified means we don't know whether it has a shell at all. If you try one, tell us what happened through the contact form.
The one-prompt way#
On the dashboard's Overview, open Set up an AI agent's computer (Devices links to it too), keep With a code and choose Make a setup prompt. You confirm it's you (passkey or email code), and it makes a one-time attach code and a prompt that contains it. Paste the prompt into the agent. The prompt tells it to install the CLI and run one command:
npm install -g 0bridge@latest
0b setup --agent-vm --platform muse --attach vm_XXXX --no-waitThe code works once and expires in 15 minutes. Because you made it after confirming it's you (passkey, or an email code if you have none), the computer signs in without you approving anything else. The landing page has the same prompt without a code, for the version where you approve a link.
Without a code: approve a link#
Without --attach, the command prints a sign-in link on its own line, then a code. Open the link on your phone or laptop: the page says an AI agent's computer is signing in. Check that the code matches, and approve. The session that approval hands over can only make the computer's expiring token. If you add your account's email (--email you@example.com, or ZEROB_EMAIL), the page also names the machine, and the request shows up on your dashboard to approve with a number match: Approvals and sign-in.
An agent's shell often times out before you get to the link. So the command doesn't wait forever: with --no-wait, or after about 9 minutes, it saves the pending sign-in, tells you to approve the link, and exits normally. Run the same command again and it picks the sign-in up. --qr qr.png also saves the link as a QR image, for an agent that can show you pictures.
What the command does#
It never asks a question, so an agent can run it. Each step prints one line, and a failure in a later step is reported while the rest carry on.
- Names the computer: the name you gave the setup code, else
--name, else the platform and the host name. - Signs in with the code or the link. The token it gets is labeled an agent computer, expires in 7 days (
--days, 1 to 30) and can be revoked. - Adds 0bridge to the Claude Code and Codex configs it finds, with your skills and instructions (AGENTS.md).
- Pulls your personal files when the folder is a linked repo and the vault is already open there; otherwise it says to run
0b files pullonce you've approved the vault request. - Turns on history and the hooks that keep your session board current.
- Asks for your vault key, unless you pass
--no-vault. It prints an approval link; nothing opens until you approve it. - Prints a summary and how to sign the agents in.
It does not install the agent daemon or any background service, and it never signs in to Claude or Codex for you. Run 0b background on yourself if the computer stays up and you want history to sync on a timer; the hooks keep it current otherwise.
Running it again is safe. Attaching under the same name again revokes the older token for that name. A run on a computer that is already signed in leaves its token alone: to renew one, run 0b logout there, then the command again with a new code.
Claude and ChatGPT sign-ins on someone else's computer#
Claude Pro and Max sign-ins are for Claude Code and Claude themselves (since 2026-02-19). For an agent working on someone else's computer, use API keys: keep ANTHROPIC_API_KEY or OPENAI_API_KEY in your vault and run the agent through 0b exec, or sign Codex in with codex login --device-auth for a ChatGPT plan. Never paste a key or token into the agent's chat.
0b secret set ANTHROPIC_API_KEY --global
0b exec -- claudeYour vault on a computer you don't own#
An agent computer belongs to its vendor's platform. Anything running on it can use every secret your vault opens there. That is why the vault request is a separate approval with a warning, and why --no-vault exists. For the same reason, its token can't start or answer agents on your other machines, change your webhooks or change your settings. The one thing it may send your machines is work for a machine in ledger mode, with 0b host (From a terminal): there everything it sends is an entry in the work ledger, and nothing is typed into a terminal. Approve it only for a computer you are willing to trust with those secrets, and prefer a small set: put only what the agent needs in one repo's own secrets.
Revoke#
Agent computers lists each one with its name, when it expires and when it was last used. Revoke removes its token at once and disconnects its agent connection, if it had one. The token also stops by itself after its days run out.