github-mcp
MCP server exposing GitHub tools for issues, pull requests, and code browsing via the GitHub REST API. Designed for local LLM clients with flat arguments and streamable HTTP support.
README
github-mcp
Thin FastMCP server exposing flat GitHub
tools (issues, PRs, code browsing) over the GitHub REST API. Built for local
LLM clients: scalar arguments, comma-separated lists instead of JSON arrays,
trimmed JSON-string responses, and failures isolated as {"error": ...}.
Why not the official github/github-mcp-server? Its default toolset alone
registers 46 tools (~10-15k tokens of schema) with rich nested arguments,
and self-hosted it is stdio-only. This server keeps 15 flat tools (~2-3k
tokens) and speaks streamable HTTP natively, which Open WebUI consumes
directly. OpenCode terminal sessions continue to use the gh CLI.
Tools
| Tool | Args | Notes |
|---|---|---|
get_me |
Authenticated user | |
search_repositories |
query, language, min_stars, page, per_page | Sorted by stars desc; repo exploration by popularity |
get_repo |
repo | Summary card: stars, forks, language, topics, license |
list_issues |
repo, state, labels, page, per_page | PRs filtered out; labels comma-separated |
search_issues |
query, repo, per_page | GitHub search syntax; repo adds repo: qualifier |
get_issue |
repo, number, include_comments | Body + up to 50 comments |
create_issue |
repo, title, body, labels, assignees | Labels/assignees comma-separated |
update_issue |
repo, number, title, body, state | state: open/closed |
add_issue_comment |
repo, number, body | |
list_pull_requests |
repo, state, page, per_page | |
get_pull_request |
repo, number | Body, first 100 files, CI check summary |
create_pull_request |
repo, title, head, base, body, draft | base defaults to repo default branch |
add_pr_comment |
repo, number, body | General comment, not a code review |
get_file_contents |
repo, path, ref | Raw text, refused above 100k chars |
search_code |
query, repo, per_page | GitHub code search syntax |
repo arguments are always "owner/repo".
Auth
One fine-grained PAT in the environment (GITHUB_PERSONAL_ACCESS_TOKEN):
Issues r/w, Pull requests r/w, Contents r (also covers releases),
Checks r (CI summary in get_pull_request; without it the summary
degrades to a checks_error hint), Metadata r (auto-granted). Actions r
is not needed today, only if workflow-run tools are added later. A local
.env file is loaded if present.
Run
# stdio (default, local MCP clients)
GITHUB_PERSONAL_ACCESS_TOKEN=... uv run github-mcp
# streamable-http (Docker does this via MCP_TRANSPORT=http)
MCP_TRANSPORT=http MCP_PORT=3102 uv run github-mcp
curl -s http://localhost:3102/mcp # -> 406 (streamable-http requires MCP headers)
Docker image: ghcr.io/awernick/github-mcp:main, published on every push to
main. Deployed on robusto by awernick/picolino-config
(robusto/github-mcp/, port 3102); Open WebUI registers it as an external
tool server (Type: MCP Streamable HTTP) at http://github-mcp:3102/mcp.
Development
uv sync --dev # or: pip install -e ".[dev]"
uv run pytest
uv run ruff check .
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。