MCP Integration

Using CorgReview as an MCP server for deep agent integration.

Overview

CorgReview includes a built-in Model Context Protocol (MCP) server. This lets AI agents call CorgReview as a tool rather than running a shell command — providing richer parameter passing and structured responses.

MCP mode is supported for Claude Code, Cursor, and Windsurf.

Setup

The easiest way to set up MCP is with corg init --mcp:

corg init --agent claude --mcp

This adds the MCP server to your agent's configuration automatically.

Manual configuration

If you prefer to configure manually, add CorgReview to your MCP settings:

{
  "mcpServers": {
    "corgreview": {
      "command": "corg",
      "args": ["mcp"]
    }
  }
}

For Claude Code, this goes in ~/.claude/settings.json (global) or .claude/settings.local.json (project-level).

How it works

Your AI agent launches corg mcp as a subprocess
CorgReview reads JSON-RPC requests from stdin
When request_human_review is called, CorgReview starts the HTTP server and opens your browser
You complete the review in the browser UI
CorgReview returns the result via stdout to your agent

Tool: request_human_review

The MCP server exposes a single tool:

Parameters

Parameter	Type	Description
commit_range	string	Git range to review. Default: auto-detect branch or HEAD~1..HEAD
repo_path	string	Repository path. Default: current working directory
focus_areas	string[]	Hints for the reviewer about what to focus on
blocking	boolean	Wait for review completion. Default: true
description	string	Branch overview/purpose shown in the review UI
agent	string	Agent name to record in review metadata

Response

{
  "status": "changes_requested",
  "review_path": ".corg/reviews/2026-03-feat-auth-flow.md",
  "review_file_path": ".corg/reviews/2026-03-feat-auth-flow.md",
  "summary": "3 comments total: 1 blocker, 1 question, 1 suggestion",
  "blocker_count": 1,
  "question_count": 1,
  "suggestion_count": 1,
  "reviewer_summary": "Overall looks good but error handling needs work."
}

Protocol details

The MCP server uses JSON-RPC 2.0 over stdio and supports these methods:

initialize — Protocol handshake
initialized — Client ready notification
tools/list — List available tools
tools/call — Execute a tool

CLI mode vs MCP mode

Both modes achieve the same result — the main differences are:

	CLI Mode	MCP Mode
Setup	Skill file only	Skill file + MCP config
Parameters	CLI flags	Structured JSON
Response	JSON on stdout	Structured MCP response
Agent support	All agents	Claude Code, Cursor, Windsurf

For most users, CLI mode is simpler and works everywhere. MCP mode provides richer integration for agents that support it.