Documentation
Agents and MCP
Connect Claude Code, Codex, or Cursor without handing approval to the protocol.
Automatic setup
rationale init --skip-agent-config
rationale install-agent --dry-run
rationale install-agent
The installer detects supported agents, writes an idempotent instruction block, and registers the MCP server where the agent supports a project configuration. The block contains the master prompt.
Claude Code also receives six project skills:
/rationale-preflight, /rationale-explain, /rationale-capture,
/rationale-health, /rationale-protocol, and the human-only
/rationale-review.
Manual setup
For a global Codex registration:
codex mcp add rationale -- "$HOME/.local/bin/rationale" serve
Codex reads the Rationale protocol from AGENTS.md. Use plain-language requests
instead of assuming a client-specific slash command, for example:
Prepare this change with Rationale for
<target>with intent<intent>.
Inspect agent configuration before committing it. Never commit tokens, private keys, or sensitive paths.
What MCP does not do
MCP exposes health, preparation, explanation, and capture. Approval, dispute, revocation, supersession, and authority changes stay in the interactive CLI.