Code Review Agent MCP

AI code review agent that doesn't sugarcoat. If your code is bad, it tells you. Bluntly.

An AI-powered MCP (Model Context Protocol) server that reviews your code the way experienced maintainers review patches on serious projects: technically, directly, without diplomacy.

What it does

• AI code review — reviews code snippets, files, git diffs, and commits

• Returns findings with severity labels: CRITICAL, MAJOR, MINOR, NIT, CLEAN

• Cites specific line numbers for every finding

• Explains why something is bad, then how to fix it

• Ends with a clear verdict: "Ship it." or "Do not merge until..."

• No false praise. No hedging. No apologies. No filler.

Related MCP server: greybeard

Why?

Most AI code reviewers are trained to be polite. They preface criticism with "great work but...", they hedge with "you might want to consider...", they soften findings with "this could potentially be an issue".

That's not how serious code review works. On real projects (Linux kernel, PostgreSQL, Redis, SQLite), maintainers review patches directly. They find problems and call them out clearly. They don't exist to make the author feel good — they exist to make the code better.

This MCP encodes that tradition. The personality IS the product.

Anti-RLHF engineering

LLMs are RLHF-trained to be polite. This AI agent fights that with three layers:

1. System prompt — 12 core principles (code-not-coder, verdict-first, severity labels, line citations, no false praise, no hedging, no apologies, no filler, second person, no demographic references, no refusal, no hallucinated bugs)

2. Post-processor — strips 40+ banned phrases (hedging, softeners, apologies, filler) from LLM output

3. Validator — checks output has severity labels, line citations, and verdict; flags vague problem statements as hallucination signals

Tools

Harshness levels

Installation

pip install code-review-agent-mcp

Or with uv:

uv pip install code-review-agent-mcp

Configuration

Claude Desktop

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "code-review-agent": {
      "command": "python",
      "args": ["-m", "blunt_codereview.server"]
    }
  }
}

Or if installed via pip:

{
  "mcpServers": {
    "code-review-agent": {
      "command": "blunt-codereview-mcp"
    }
  }
}

Cursor

Add to .cursor/mcp.json:

{
  "mcpServers": {
    "code-review-agent": {
      "command": "python",
      "args": ["-m", "blunt_codereview.server"]
    }
  }
}

Usage examples

Review a code snippet

User: Review this code for me

def get_user(username):
    import sqlite3
    conn = sqlite3.connect("users.db")
    cursor = conn.cursor()
    query = f"SELECT * FROM users WHERE username = '{username}'"
    cursor.execute(query)
    return cursor.fetchone()

MCP response:

## Code Review: snippet

### Findings

**CRITICAL** `snippet:7` — SQL injection
The query uses an f-string with user input, allowing SQL injection. Use parameterized queries: `cursor.execute("SELECT * FROM users WHERE username = ?", (username,))`.

### Verdict

Do not merge until CRITICAL is fixed.

Review a file

User: Review src/auth.py

MCP calls review_file with file_path="src/auth.py"
Returns blunt review with line citations.

Review a commit

User: Review the last commit

MCP calls review_commit with commit_ref="HEAD"
Returns blunt review of the diff.

Severity labels

Security

This MCP server implements security sandboxing:

• File access is sandboxed to the current working directory by default

• Sensitive paths (.ssh, .aws, .env, /etc/passwd, etc.) are refused

• Git refs are validated against a strict character whitelist to prevent option injection

• Subprocess calls use shell=False and disable global git config

See SECURITY.md for the full threat model.

Development

# Install in development mode
pip install -e ".[dev]"

# Run tests
pytest

# Run tests with coverage
pytest --cov=blunt_codereview

Benchmark snippets

The benchmarks/ directory contains 5 regression snippets that verify the reviewer:

1. SQL injection — expects CRITICAL, line citation, "Do not merge"

2. Mutable default — expects MAJOR

3. Clean code (binary search) — expects CLEAN, "Ship it" (anti-hallucination test)

4. Swallowed exception — expects MAJOR

5. Off-by-one — expects MAJOR

The clean code benchmark is the most important — it catches hallucination. If the reviewer invents bugs in correct code, the anti-RLHF system is broken.

License

MIT

Acknowledgments

This project encodes the kernel maintainer tradition of code review — a methodology practiced by many senior engineers across many projects (Linux kernel, PostgreSQL, Redis, SQLite, and others). We cite the tradition, not any single practitioner.