Documentation

MCP reference

Tool-by-tool behavior at the agent boundary, including what MCP can never approve.

`health`

Returns project identity, Git revision, provider status, and coverage. A missing Codebase Memory provider is an explicit degraded result, not a reason to invent symbol resolution or block the change.

`prepare_change`

Input: target and the agent’s actual intent; severity is explicit when a new assertion is being captured. Output includes constraints, evidence, authority, linkage, provider coverage, intent conflicts, and whether a governance verdict is required.

Governing Records are not hidden because their severity is medium, and an empty match is honest: the server never falls back to the first unrelated Record.

`explain_target`

Returns the same governing Record set for a target without an intent. It uses the same binding matcher as prepare_change, so the two tools do not disagree about a file-only binding or a structural suffix.

`finalize_change`

Captures mechanically observed files, provider resolution, intent, evidence, and a proposed statement. Uncommitted files are marked provisional and remain capturable; a proposal is still pending until a human reviews it.

Prompts

The server advertises the MCP prompts capability. prompts/list returns six pre-made actions from the same source used to generate Claude Code skills:

Prompt	Purpose	Arguments
`preflight`	Prepare constraints and intent conflicts before editing.	`target`, `intent`
`explain`	Explain a possible Chesterton fence before simplifying it.	`target`
`capture`	Guide `finalize_change` after a change.	optional `statement`
`review`	List pending proposals and hand approval to the human CLI.	none
`health`	Diagnose MCP, provider, Git, and canon health.	none
`protocol`	Load the complete master protocol.	none

prompts/get substitutes named arguments and returns one user message. An unknown prompt is a JSON-RPC error and does not terminate the persistent session. Prompt discovery and command decoration belong to each MCP client; do not rely on a slash-command name without verifying that client. In Codex, ask in plain language, for example: “Prepare this change with Rationale for <target> with intent <intent>.” Claude Code separately receives the clean /rationale-preflight project skill.

Boundary

MCP does not expose approval, correction, dispute, revocation, supersession, or authority changes. Those actions require the CLI’s interactive review path. The stdio protocol is JSON per line; Content-Length is used only when the client talks to Codebase Memory.