Agents

Agents on your machine

This is the machine path: create an agent and give it a policy, connect it to your machine (through Claude Code, a runner, or both), and use it from inside a topic. A few minutes, and you only connect once per machine.

Cloud agents skip all of this

A cloud agent has no machine to pair — the owner connects GitHub once for the organization and a topic is pointed at an allowed repository. See Cloud agents.

1 · Create an agent

Open Agents & MCP (the laptop icon in the top bar) → New agent, choose My machine, and give it a name and what it’s for. Then, on its card, set its standing policy — this is what keeps it under your control no matter who asks it for work:

Repos it may work in — tick the repos your machines offer, and pick a capability for each: Read only, Edit files, or Edit and run.
Who can invoke it — Only me, Anyone in the topic (the default), or Anyone in the organization. Anyone outside that circle waits for your approval.
How it should behave — plain instructions, e.g. “Be terse. Say what you changed and why.”

An agent belongs to one org

It’s created inside the organization you’re in and is invisible everywhere else. See What are agents for the principles.

2 · Connect it to your machine

Pair a machine so the agent can work while you’re away, when someone mentions it in a topic. Connecting your own editor to your to-dos is a separate step — see MCP.

The runner — pair a machine for your agent

On the Agents & MCP page, Your machines → Connect a machine walks you through it in about three minutes.

Install the runner (first time only):

$curl -fsSL https://<your-workspace>/install | sh

Pair it with the one-time command shown there (valid 30 minutes), naming the repos this machine may be used for:

$teammate-runner pair tm-XXXXXX --repo ~/code/app --repo ~/code/api

Keep it running, so it starts at login and restarts if it crashes:

$teammate-runner install

The machine sets the maximum

Only repos named with --repo (or teammate-runner repos --add) can be offered. The policy on the website can narrow that list — never widen it.

3 · Use it

From inside a topic, @mention the agent by its @handle or hand it a to-do. If you (or a pre-approved teammate) asked it, it runs; otherwise it waits for your approval. When someone new asks, a card appears under Waiting on you on the Agents & MCP page, and the laptop icon in the top bar shows a badge:

Button	Effect
Approve once	Runs this request.
Always allow <name>	Approves it, and future requests from this person at this capability run without asking.
Edit task	Rewrite the instruction before approving.
Decline	Requires a reason, which they’ll see.

Once approved, the task is claimed, run in the repo, and reported back into the topic — a live status line (working · 0:42 · Read 6 files · running npm test) becomes the final answer and diff, with the files changed, time and cost.

Safety guarantees

Clean-tree check — it won’t edit a repo that has uncommitted work of yours.
Changes stay local — edits are left uncommitted for you to review; the agent never commits or pushes on its own.
Memory — follow-ups in the same topic and repo continue the same Claude session.

Runner commands

Manage a paired machine with teammate-runner status, repos --add / --remove, start / stop, upgrade and uninstall.