A cog runs a coding agent on one of your computers and produces a reviewable artifact. Fill in the task and pick a computer — everything else is set to safe defaults you can ignore.
The taskrequired
The prompt the coding agent runs. Describe one job with a clear, verifiable output.
Auto-suggested from your task; editable.
HarnessClaude Code
The coding agent the cog runs. It must be installed on the target computer (e.g. the "Claude Code dev box" app); the cog run invokes it headless. Harnesses marked "coming soon" are not selectable yet.
Where it runsrequired
Only running computers you own are listed.
/config/cogs/<cog_id>
The cog's slice of the computer's filesystem (assigned on create; shown, not editable).
Tools this cog may usegranted: 0 of 0
Nothing is granted unless you check it. A cog with no tools runs with shell and files only.
Loading connected tools…
No tools are available to grant yet. Install a tool-providing app (such as coordd) from the App Store, then grant it here.
Credentials never appear here — this grants access to a tool, it never handles the tool's secret.
Triggermanual
Run it yourself first; add a schedule after the first reviewed run.
5-field cron. The preset writes this; edit it directly for anything custom.
A token is generated on create and shown once. Whatever the caller sends arrives marked untrusted — your task decides whether to use it.
Outputreport
What the run produces for you to review.
Defaults to a pull request when a repo is set, otherwise a report file.
Autonomysafe defaults
PR-not-push · first run is a dry run · $5.00/run cap · 1h timeout · approval gate on