How it works

A summons is a row

When you assign your agent a ticket or mention it, the app writes a summons: which agent, what it’s about (a ticket key and its URL), why (assigned, mentioned), and the repository and branch if there is one. It never holds the words that asked. The agent reads those over its own connection to the app, where its access is checked at that moment.

A place claims it

A place is somewhere your agent can start: Wake on one of your machines, paired to that agent. Wake keeps one Convex subscription open per app. When a summons appears, every place of that agent sees it at once, and each asks to claim it. The first claim wins; the rest let it go. A Mac that was asleep claims what’s still open when it wakes.

Only you start runs unattended

A summons you make starts right away. One someone else makes (a teammate assigning your agent) waits for your yes, because a run writes code on your machine.

The run

The winning claim hands Wake the app’s instructions: which tools to call to read the ticket, and how to finish. Wake finds your clone, makes a worktree on the ticket’s branch, and starts the agent there, fenced (see Security).

As the agent starts, Wake tells the app it’s on it, under the run’s name, two words like amber-heron, with the tool and its session. The app shows that where you asked (“Claude is on it as amber-heron”) instead of the agent writing a comment to say so, and shows the machine and the session to you alone. An app that doesn’t take the report gets the agent’s comment instead.

The agent does the work, opens a pull request if the ticket asked for code, and comments what it did. Wake then tells the app the run is done, or failed and why. Apps can also list your agents’ runs outside the comments, in their top bar.

If nothing claims a summons in thirty minutes, the app says so where it was made. It stays open: your Mac will still claim it when it’s back.