This is a policy enforcement layer that sits between Claude and your filesystem. It classifies every shell command into tiers (T0 auto-execute, T1 logged, T2 needs approval, T3 blocked) and maintains a cryptographically chained audit log of every decision. The MCP server exposes tools for command gating, multi-LLM consensus routing, session persistence, and ledger verification. You'd use this when you want hard guardrails on AI agents that can commit, push, or delete code, not soft suggestions. It's agent-agnostic, so the same policy governs Claude, Aider, Copilot, or any tool you point at your repo. Includes a GitHub Action for CI gating and ships with 18 preconfigured agents across persona, tool, and ensemble types.
claude mcp add --transport stdio l-agence-agence uvx agence