Run a saved task automatically
Pair a runner, test a profile, and start after recipient consent.
Manual Docked sessions use your connected agent. Automatic sessions use a separate supervised runner that starts the saved task when the recipient clicks Start. The original chat does not need to stay open.
Set up a runner
- Open Settings → Connections → Add runner. Choose a local computer or an always-on server you manage, plus the runtime and concurrency.
- Configure the runner from the Docked source checkout using
runner/deploy/config.example.json. Set an absolute private state directory, pinned runtime path/version, provider account reference, and isolated credential file. Do not put provider keys into the dashboard. - Run
node runner/cli.mjs inspect /absolute/path/to/config.json. Resolve any installation or isolation errors. - Create a pairing code in Connections and run the displayed command. The one-use code expires in 10 minutes.
- Run
node runner/cli.mjs run /absolute/path/to/config.json, or use the documented launchd/systemd service. Local runners require the computer to stay awake. - Add an execution profile with the exact runtime, model, account reference and limits. Save & test, then activate the validated revision.
The deployment must have durable Redis storage, a healthy external reconciler, and a release-certified runtime adapter before automatic execution can be enabled. An installed runtime alone is not a passing certification.
Create a session
Choose an active profile under Run with. The recipient sees the task, permissions and automatic-start disclosure before consenting. The invitation stays pinned to that configuration even if you change your defaults later.
When a run stops
Pause stops new browser actions. A known, settled process can continue its exact saved conversation only on an adapter certified for resume. An uncertain action or unknown process result requires reconciliation; Docked does not replay it. Cancel run revokes the attempt. Take over manually transfers browser control after fencing the automatic agent.
Refresh and revoke
Refresh checks the saved configuration without replaying the task. Drain prevents new jobs while existing leases settle. Revoke stops the runner's execution authority. Replacing a runner creates a new pairing; it cannot inherit an uncertain process from another machine.
At this release, all four adapter implementations have deterministic contract tests. Live runtime certification, always-on acceptance and the soak gate must pass separately before production enablement. Codex and Claude Code have exact-conversation continuation paths; Hermes and OpenClaw continuation remains blocked until certified.