Recorded from the live app: describing a job, watching the name auto-suggest, and the autonomy card — PR-not-push, dry-run first, a per-run cap, approval gate.
Describe the job once; it runs on a schedule
A cog is one job with a clear, verifiable output — not an always-on agent loop. You review the first run, then let it repeat.
Write one job in plain language — “review open PRs and comment on each.” The name auto-suggests; everything else has a safe default you can ignore.
Choose the coding-agent harness (e.g. Claude Code) and the computer it runs on. The cog invokes the harness headless on your box.
Run it yourself first — the first run is a dry run. Once you have reviewed a clean run, add a schedule and it repeats.
Autonomy with a seatbelt
Safe defaults you can ignore
Every cog ships with the guardrails on: PR-not-push, the first run is a dry run, a per-run cost cap, a one-hour timeout, and an approval gate. A cog with no tools granted runs with shell and files only. You opt into more, never out of safety.
/cogs/new
Honest status, not a spinner forever
The runs view tells you the truth about each run — queued, running, needs review, done, or failed. A cog produces a reviewable report you actually read before anything ships. No mystery, no runaway loop.
/cogs
Start from a recipe
Don't want a blank form? Start a cog from a workflow template — a repeatable recipe of steps rather than a fragile one-off script. Tweak a couple of fields and run it on any of your computers.
/workflowsVerified by browser tests
The cog create flow and runs view are covered by a green Playwright end-to-end suite against the live app.
Real screenshots
Everything on this page is captured from the running application — the actual form, the actual list.
Cost-capped by design
A per-run cap and a monthly budget mean a scheduled agent can never surprise you with a runaway bill.
Schedule your first coding agent
Describe one job, pick a box, review the first run — then let it work every morning.
Start on your box