What the agents actually do

Each agent has a role, the access you give it, and work it keeps without being asked. One page per job, with what it does, what you have to grant it, and where it stops.

• Answering customer email

  Your mail, in your words, with the ones it should not answer by itself passed to you.

• Watching the money

  A refund, a failed payment, a complaint: it checks what happened, acts inside the limit you set, and asks when it is outside it.

• Fixing the bug

  It reads the repository you gave it, pushes the fix, runs your own checks, and tells you what it ran.

• Shipping the small thing

  The feature you never get to. A commit, a deploy, and a note saying what changed.

• Keeping watch

  Work it keeps without being asked: it checks on its own schedule, and what it finds it sees through to the end.

What is the same on every one

The agent runs on a server you own and signs in to the Claude plan you already pay for. Its access is granted one grant at a time, from a vault on your machine or a GitHub App in your own organisation. Every grant and every owner decision is a commit you can read and revert, and anything it was not granted reaches your phone as a question rather than as a surprise.

One subscription covers one fleet — a hub, its machines and its agents — with fair use of 5 machines and 10 agents, so these are jobs to run together rather than one at a time. What that costs is one page.

Where to start

With one agent and one job, whichever one is costing you the most attention. Watch what it spends and what it gets right, and add the rest when it has earned it. Before any of them, decide what it may reach — that is the decision this product cannot make for you.