claw-drive →
Drive Claude Code. End-to-end.
Policy-gated. Human-supervised. Restart-resilient.
curl -fsSL https://raw.githubusercontent.com/renatodarrigo/claw-drive/main/install.sh | bashHow it works
Session A hosts claw-drive as an MCP server and can spawn one or more driven sessions — Session B above, plus a Session C, D, … by calling start_session again. Each is a real claude -p subprocess in its own directory, with its own policy, scenario brief, and event log; they run in parallel. Every tool call a driven session makes is gated by a policy that can approve, reject, defer to the human, or escalate to A.
What you get
Policy gates
auto_approve, auto_reject, auto_defer, escalate_default. Rule-shape learning via remember_as_policy. Two templates ship — conservative and permissive.
Human-in-the-loop
Escalated calls surface to you. Deferred calls — sudo, interactive logins — round-trip through provide_tool_output. No silent bypasses.
Monitor-driven
Event-stream driving via watch_command, not polling loops. A reacts to tool_decision_required, turn_completed, and three other kinds.
Deeper reading
Install →
One curl, two Bash deps. Under a minute.
Plugin →
Claude Code slash commands: /claw-drive-init, /claw-drive-start, /claw-drive-resolve.
Policies →
The full policy model: templates, evaluation order, CLAW-GATE, remembering decisions.
Driving patterns →
Monitor + watch. Defer round-trip. CLAW-GATE. Interrupts. Restart resilience.
Reference →
Every MCP tool and CLI command, in two tables.
Contributing →
Tests, PR shape, where to file bugs.