git-mcp
Connects Claude to GitHub via OAuth and proxies official GitHub MCP tools, allowing each end user to interact with GitHub as themselves.
README
git-mcp
A remote MCP server that connects Claude to GitHub on behalf of each end user.
- Auth — runs Claude's custom-connector OAuth flow against a GitHub OAuth App
(FastMCP
GitHubProvider). Each user logs into GitHub through Claude. - Proxy — proxies the official
github/github-mcp-server(run inhttpmode), forwarding each request with the authenticated user's GitHub token. All of GitHub's official MCP tools are exposed to Claude, executed as that user.
Claude custom connector ──HTTPS /mcp──▶ FastMCPProxy (this app)
auth = GitHubProvider (OAuth App) → per-user GitHub token (get_access_token().token)
client_factory (per request) → Authorization: Bearer <user token>
▼
github-mcp-server (official, http mode, 127.0.0.1:8082)
▼
GitHub API (as the user)
Stack
Python 3.12+ · uv · ruff · ty · pytest · fastmcp (pinned).
Setup
make install-dev # uv sync + pre-commit hooks
cp .env.example .env # then fill in the values
Create the GitHub OAuth App
GitHub → Settings → Developer settings → OAuth Apps → New OAuth App.
- Homepage URL: your
BASE_URL(e.g.https://your-app.herokuapp.com) - Authorization callback URL:
<BASE_URL>/auth/callback
Copy the Client ID / secret into .env (GITHUB_OAUTH_CLIENT_ID, GITHUB_OAUTH_CLIENT_SECRET).
Develop
make lint # ruff check + format check + ty
make test # pytest -n auto
make format # auto-fix + format + ty
Run locally
The app spawns github-mcp-server http as a subprocess, so you need the binary on
PATH (or set GITHUB_MCP_BINARY). Easiest is a local build/binary from the
releases; alternatively run the
backend yourself in Docker via make run-backend and point BACKEND_MCP_URL at it
(then comment out the subprocess spawn for local testing).
make run-dev # binds 0.0.0.0:$PORT (default 8000), serves /mcp
Test the OAuth + tool flow with the MCP Inspector:
npx @modelcontextprotocol/inspector
# connect to http://localhost:8000/mcp → completes GitHub OAuth → lists tools
Deploy to Heroku (container)
Both processes share localhost, so they run in one dyno via a multi-stage image.
heroku create your-app
heroku stack:set container -a your-app
heroku config:set -a your-app \
GITHUB_OAUTH_CLIENT_ID=... \
GITHUB_OAUTH_CLIENT_SECRET=... \
BASE_URL=https://your-app.herokuapp.com \
JWT_SIGNING_KEY="$(openssl rand -hex 32)" \
GITHUB_TOOLSETS=repos,issues,pull_requests,users,context \
GITHUB_SCOPES="repo read:org read:user"
git push heroku main
BASE_URL and the OAuth App callback (<BASE_URL>/auth/callback) must match the
deployed URL exactly.
Add to Claude
Add https://your-app.herokuapp.com/mcp as a custom connector (Streamable HTTP).
Claude runs the OAuth flow; after you authorize the GitHub OAuth App, the GitHub tools
become available, acting as your GitHub user.
Pinning
Keep both moving pieces pinned:
fastmcp(proxy/auth API moves across versions; targets v3.4.2+).ghcr.io/github/github-mcp-serverimage (pin a digest in theDockerfile).
Configuration
See .env.example. Notable: GITHUB_TOOLSETS (or all), READ_ONLY=1 to force
read-only tools, GITHUB_SCOPES for the OAuth scopes the proxied tools need.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。