Your agents plan.
You approve.
Chief keeps the record.
Agents improvise around broken steps, bury your feedback in chat scrollback, and report “done” with nothing to show for it. Chief makes the plan a contract: it waits for your approval before it can run, mid-run changes pause for your decision, and every transition lands in an audit log.
Chief never executes anything. It records what a harness plans and what it reports, and it enforces who has to say yes.
Claude Code — or anything that can make an HTTP call — plans a workflow: one step per unit of work, ordered by explicit dependency edges. It arrives as a draft.
A draft cannot take a run until you approve it. Leave review notes on any node; the agent reads them off the plan it fetches before revising. Nothing to repeat.
When the plan stops fitting, the agent proposes an amendment and the run pauses until you decide. Touching finished work is a history edit — never auto-approved.
Running in under a minute
Needs Python 3.11+ and git. One process serves the REST API, the web UI and the MCP endpoint. No build step, no CDN, no separate UI command — and the database is one SQLite file you back up by copying.
Then point Claude Code at it:
claude mcp add --transport http chief http://127.0.0.1:8080/mcp/ -s user
$ git clone https://github.com/zeroshotmind/chief.git $ cd chief && python3 -m venv .venv && source .venv/bin/activate $ pip install -e ".[dev]" $ chief --port 8080 → serving UI, API and MCP at http://127.0.0.1:8080/
Agents move fast. Keep the record straight.
If human-approved agent workflows are something you want to exist, a star is how you say so.
★ Star zeroshotmind/chief local-first · no auth, no cloud, no telemetry