tai-mcp-github
# tai-mcp-github
A minimal [MCP (Model Context Protocol)](https://modelcontextprotocol.io) server that exposes a focused set of GitHub PR review tools to AI agents.
## Why
The official `github/github-mcp-server` exposes 60+ tools (Actions, Discussions, Dependabot, secret scanning, gists, etc.). This inflates the MCP schema sent to the LLM on every request, increasing token usage and noise. This server limits the surface to exactly what is needed for a PR review workflow — nothing more.
## Tools
| Tool | Description |
|---|---|
| `list_pull_requests` | List open (or closed/all) PRs for a repository |
| `get_pull_request` | Get full details of a specific PR |
| `list_pr_comments` | List general thread comments on a PR (conversation tab) |
| `list_unresolved_review_threads` | List unresolved inline review threads (diff comments) |
| `resolve_review_thread` | Mark an inline review thread as resolved |
## Requirements
- Python 3.12+
- [uv](https://docs.astral.sh/uv/)
- A GitHub Personal Access Token with **Pull requests: Read** (and **Write** if you want to resolve threads)
## Setup
```bash
git clone <repo-url>
cd tai-mcp-github
cp .env.example .env
# Edit .env and set GITHUB_TOKEN=your_pat_here
```
## Usage
### Run the server
```bash
uv run server.py
```
### Use with Claude Code
Add to your MCP configuration (e.g. `~/Library/Application Support/Claude/claude_desktop_config.json`):
```json
{
"mcpServers": {
"github-pr-review": {
"command": "uv",
"args": ["run", "server.py"],
"cwd": "/path/to/tai-mcp-github"
}
}
}
```
## Development
### Run tests
```bash
uv run --with pytest --with pytest-asyncio pytest tests/ -v
```
### Lint and type-check
```bash
uv run --with ruff ruff check .
uv run --with ruff ruff format --check .
uv run --with mypy mypy server.py
```
### Git hooks
Pre-commit (ruff + mypy) and pre-push (tests) hooks are provided. Install them once:
```bash
cp .git/hooks/pre-commit .git/hooks/pre-commit # already in place after clone
```
> Note: git hooks are local only. Run the script below on a fresh clone to activate them.
```bash
chmod +x .git/hooks/pre-commit .git/hooks/pre-push
```
## Tech stack
- **[fastmcp](https://github.com/jlowin/fastmcp)** — MCP server framework
- **[PyGithub](https://pygithub.readthedocs.io)** — GitHub REST API client
- **[httpx](https://www.python-httpx.org)** — HTTP client used for GitHub GraphQL API calls
- **[python-dotenv](https://saurabh-kumar.com/python-dotenv/)** — `.env` file support
- **[ruff](https://docs.astral.sh/ruff/)** — linting and formatting
- **[mypy](https://mypy-lang.org)** — static type checking
- **[pytest](https://pytest.org)** — testing
## Architecture
```
server.py # Single-file MCP server — all tools live here
tests/
└── test_server.py # Unit tests (fully mocked, no GitHub calls)
```
The server is intentionally stateless — no caching, sessions, or background tasks. Each tool maps directly to one GitHub API call:
- PR listing and detail use the **GitHub REST API** via PyGithub
- Review thread listing and resolving use the **GitHub GraphQL API** via a single `httpx` POST (the GraphQL `resolveReviewThread` mutation is not available on the REST API)
Authentication is via a GitHub Personal Access Token passed through the `GITHUB_TOKEN` environment variable. No OAuth or GitHub App flows are needed.
## Scope
This server is read-focused. Explicitly out of scope:
- Git operations (commits, pushes, branch management) — do these locally
- Any GitHub feature outside the PR review workflow (Actions, Issues, Releases, etc.)
TDQS
Scored across 5 tools
Each tool targets a distinct aspect of pull request workflows: listing PRs, getting a specific PR, listing comments, listing unresolved review threads, and resolving them. There is no overlap or ambiguity.
All tool names follow a consistent verb_noun pattern (list_pull_requests, get_pull_request, list_pr_comments, list_unresolved_review_threads, resolve_review_thread), using lowercase with underscores.
Five tools is an appropriate and focused scope for a GitHub pull request review assistant. Each tool earns its place, covering listing, detail, comment viewing, unresolved thread viewing, and resolution.
The tool set covers the core PR review workflow well—listing, reading, and resolving threads. However, it lacks tools to create or reply to comments or threads, which agents may need to fully participate in reviews.