Docs
The agent
The agent does work on your own Mac, in apps you're already signed in to, using Claude's computer use plus the accessibility tree — with the compiled playbook as its guide.
Work packages
A work package is what you hand the agent:
- Playbook and version to follow, plus optional extra instructions.
- Inputs — the playbook's parameters, such as a ticket number or amount.
- Scope — the apps and domains it may touch. Anything else is refused.
- Approval mode — commit actions require approval (default) or the playbook runs autonomously.
- Budgets — max steps (default 60) and minutes (default 20).
- Success criteria — what “done” looks like, checked in verification.
Starting a run
Use Run with agent on a playbook or the Run playbook button in the web app, Run playbook… in the Mac menu bar, or the start_run tool from Claude. Runs are delivered to your Mac over a secure WebSocket; if it's offline, the run waits in the queue.
Run states
| State | Meaning |
|---|---|
queued | Waiting for your Mac to come online and claim it. |
claimed | Your Mac picked it up and is loading the playbook. |
running | The agent is working through the steps. |
awaiting_approval | Paused before a commit action until you approve or deny. |
verifying | Checking the result against success criteria and the compiled after-conditions. |
succeeded | Done and verified. |
escalated | Stopped and handed back: the case wasn't covered, or you denied an approval. |
failed | Something went wrong; nothing beyond the logged actions was changed. |
cancelled | You stopped it (kill switch, menu bar, web or Claude). |
Approvals
Before every click or key action, your Mac checks the target element. If it matches commit patterns — Send, Submit, Pay, Refund, Delete, Approve, Merge, Post — or the compiled step is marked commit, the run pauses. Approve or deny in the overlay on your Mac, in the web app, or with approve_step from Claude.
Kill switch
Press Esc twice anywhere, or choose Stop in the menu bar. The overlay always shows what the agent is doing, for example “Agent working: refund 3/10”. Steps that need focus wait until you're idle, or ask first.
Escalation
When a case isn't covered by the playbook, the agent ends that item with escalate and tells you why. It never invents a new branch. Escalations often become proposals once you handle the case yourself.
Audit log
Every action, element hit, before/after accessibility change, model decision, approval and screenshot is recorded in an append-only audit log under Object Lock. Open any run to review it.
Earning autonomy
A playbook can be promoted to autonomous after a number of supervised runs with no failures — 5 by default, adjustable in Settings. You can demote it at any time.
Compiled replay
Each playbook is compiled into machine-readable steps: element fingerprints, parameters, before/after conditions, risk and decision rules. In v1 this is passed to the agent as context, and a shadow mode logs what replay would have done to build a readiness score per playbook. Replay never acts on its own.