← Back to Docs
MCP

A second reviewer for your agent's plan

What is the Plan Review MCP?

The Plan Review MCP is a remote server that checks an AI coding agent's implementation plan against your codebase's real conventions before any code gets written. It runs as a Model Context Protocol server, so Claude Code, Cursor, Codex, and any MCP-capable agent can call it directly during planning, catching drift, duplication, and unsafe schema changes before they exist, not after a pull request opens.

What skills can I run?

Four skills are live today. full-review runs all twelve plan-check dimensions and is the default — run it on every plan before writing code. The other three are narrower, cheaper follow-up checks for a specific question mid-task.

SkillWhen to use itCost
full-reviewReview this PRD/spec against our codebase — the default.Included
architecture-fitDoes this plan fit how our codebase is built?Included
schema-reviewIs this migration safe?Included
security-reviewWhat's the threat surface of this plan?Included

What tools does the MCP server expose?

Six plan-review tools cover that workflow: discovering which skills exist, reading a skill's full instructions, resolving project context, running a review, fetching a persisted result, and looking up a project by name. The same server also exposes the five Feature Factory cockpit tools listed below.

ToolPurpose
list_skillsList available plan-review skills plus planned ones.
get_skillFull markdown instructions for a named skill.
get_project_contextStack profile, ADR titles, module directories, and the skill registry for a project.
vibecop_checkRun a skill against a submitted plan and extracted intents; returns verdict, findings, recommendations.
get_plan_reviewFetch a persisted plan review by id.
resolve_projectLook up projects by name; returns candidates when ambiguous, never silently picks one.

What are the Feature Factory cockpit tools?

Five additional tools drive the Feature Factory from the same MCP server: create a graph-grounded feature run, read its live state, claim and report on individual task nodes, and approve or cancel the run at its human gate. Every MCP-capable agent gets identical behavior — no client is privileged.

ToolPurpose
feature_startCreate a feature run from a title + task list with expected footprints.
feature_statusLive run view: every node's state, edges, PR and issue numbers.
node_claimClaim a ready node — returns the working branch and a 60-minute lease.
node_reportReport the branch pushed and PR open — the only self-reported status.
gate_approveApprove the task graph (G1), or cancel the run.

How do I connect my agent?

Mint a token in VibeCop Settings under "Plan Review MCP · Agent Auth," then add the server to your agent's MCP config. The full installer has copy-paste setup for Claude Code, Codex, and Cursor.

Open the MCP installer
Related docs