AI Code Review Agent

An AI-powered agent that reviews a GitHub Pull Request and posts meaningful inline review comments directly on it -- triggered by an MCP tool call, so you can just prompt an MCP-connected AI agent (e.g. Claude Desktop / Claude Code) with something like:

Could you please review this PR: https://github.com/owner/repo/pull/123

No terminal command and no manual PR-link attachment step is needed once the MCP server below is connected -- the agent calls the review_pull_request tool on its own when you ask it to review a PR.

How it works

Claude / AI Agent
        |
        | MCP tool call: review_pull_request(pr_url)
        v
MCP Server (app/mcp/server.py)
        |
        v
Review Pipeline (app/review/pipeline.py)
        |
        +-- Fetch PR + changed files + diffs      (app/github/*)
        +-- Parse diffs into hunks/lines           (app/review/line_analyzer.py)
        +-- Build LLM context per hunk             (app/review/context_builder.py)
        +-- Optional static analysis (best-effort) (app/analyzers/*)
        +-- AI review via Claude                   (app/llm/*, app/review/reviewer.py)
        +-- Validate findings (drop noise)         (app/review/validator.py)
        +-- Generate comment bodies                (app/review/comment_generator.py)
        v
Post inline comments to GitHub                     (app/github/reviews.py)

The reviewer only comments where it finds a real, meaningful issue (bugs, security, performance, logic, error handling, missing validation, etc.) -- not on every changed line, and never on style or formatting nits.

Related MCP server: Narad GitHub Agent

1. Setup

cd ai-code-review-agent
python3 -m venv .venv
source .venv/bin/activate          # Windows: .venv\Scripts\activate
pip install -r requirements.txt
cp .env.example .env

Edit .env:

GITHUB_TOKEN=ghp_your_personal_access_token       # needs "repo" scope
ANTHROPIC_API_KEY=sk-ant-your-api-key

Optional static analyzers (Ruff, Bandit, MyPy) are used automatically if installed on your machine (pip install ruff bandit mypy), and are silently skipped if not -- they never block the AI review.

2. Run the tests

pip install pytest
pytest -v

3. Try it without MCP first (dry run, no posting)

from app.review.pipeline import run_review_pipeline

result = run_review_pipeline(
    "https://github.com/owner/repo/pull/123",
    post_to_github=False,   # set True to actually post comments
)
print(result)

4. Connect the MCP server to Claude

Claude Desktop -- add this to your claude_desktop_config.json (Settings -> Developer -> Edit Config):

{
  "mcpServers": {
    "ai-code-review-agent": {
      "command": "/absolute/path/to/.venv/bin/python",
      "args": ["-m", "app.mcp.server"],
      "cwd": "/absolute/path/to/ai-code-review-agent"
    }
  }
}

Restart Claude Desktop. The review_pull_request tool will then show up as an available tool.

Claude Code -- from the project directory:

claude mcp add ai-code-review-agent -- .venv/bin/python -m app.mcp.server

5. Use it

Once connected, just prompt the agent naturally:

Review this PR and add review comments: https://github.com/owner/repo/pull/123

The agent recognizes the request, calls review_pull_request(pr_url) on the MCP server, and the pipeline above runs end to end -- fetching the diff, reviewing it, and posting inline comments to the PR.

Project structure

See app/ for the implementation, organized by responsibility: github/ (GitHub API), review/ (pipeline, diff parsing, context building, validation, comment generation), llm/ (provider-agnostic LLM client + prompts), analyzers/ (optional static analysis), models/ (data structures), mcp/ (MCP server + tools), core/ (config, logging).

Notes and limitations

• The pipeline works from PR diffs via the GitHub API; it does not check out the full repository. Static analyzers therefore run against added-lines snippets only, not whole files.

• Only claude is implemented as an LLM provider today. app/llm/client.py is structured so OpenAI/Ollama/Gemini can be added as additional classes without changing the pipeline.

• A GitHub review is posted as a single review with all inline comments batched together (capped at 50 comments per review).