ghcli-mcp-connector
ghcli-mcp-connector is an MCP server that lets agents run any GitHub CLI (gh) command—repos, issues, PRs, releases, workflows, gists, and more—as a single static Go binary with a read-only-by-default safety gate on state-changing operations.
README
<div align="center">
🔌 ghcli-mcp-connector
Talk to the GitHub CLI (gh) from an MCP-speaking agent.
</div>
A single static Go binary that speaks the Model Context Protocol
and lets an agent run GitHub CLI commands: any gh <command> <subcommand>
invocation — repo, issue, pr, release, gist, workflow, run,
secret, variable, project, ruleset, codespace, extension, api,
and more — with a read-only-by-default safety gate on anything that
changes state on GitHub.
No Python, no uv, no runtime dependency to install — just a binary and
an .mcp.json. It shells out to the gh binary already installed and
authenticated on the host (gh auth login, stored in the OS keychain, or
GH_TOKEN/GITHUB_TOKEN) instead of reimplementing a GitHub API client,
so it gets the full breadth of the CLI for free rather than a hand-curated
subset of endpoints.
Note: this is a sibling of github-mcp-connector, not a replacement. That one talks to the GitHub REST API directly via
google/go-githubwith a curated set of 18 tools. This one shells out to theghCLI itself — same trade-off asaws-mcp-connectorvs. a hand-written AWS SDK client: broader coverage (gh extensions, gh's own filters/formatting, workflow/run/codespace/project/ruleset management) through a singleghcli_execescape hatch, instead of a curated surface.
✨ Why this exists
An agent that only has a narrow, hand-picked set of GitHub tools hits a wall the moment you need something outside that set. This connector instead wraps the
ghCLI itself, so an agent can rungh pr list,gh issue create,gh workflow run,gh api repos/owner/repo/issues— anything the CLI can do — without waiting on a new tool to be written for it. State-changing commands are blocked by default and require both a server-level opt-in and a per-callconfirm=true, so exploring/ reading is safe out of the box.
🧰 Tools
| Tool | What it does | Write? |
|---|---|---|
ghcli_exec |
Run any gh <command> <subcommand> ... command. Read-only by default — commands that change GitHub state need GHCLI_MCP_ALLOW_WRITE=true on the server and confirm=true on the call. |
✅ (gated) |
ghcli_help |
Show gh <command> [subcommand] --help text — always safe, use it to check exact syntax before calling ghcli_exec. |
|
ghcli_whoami |
Show the GitHub identity (login, name, profile URL) the configured auth resolves to. |
Every tool accepts an optional response_format: markdown (default,
readable for a chat UI) or json (for programmatic use).
This server has no working-directory git repo, so repo-scoped commands
need an explicit -R/--repo owner/repo rather than relying on gh's
cwd-based repo detection.
🚀 Quickstart
Fastest path: grab a prebuilt bundle from the latest release —
download ghcli-mcp-connector-plugin-<version>-<os>-<arch>.zip, unzip it,
and point Cowork/Claude at the plugin/ folder inside (see step 4 of
SETUP.md). No Go toolchain required.
From source:
# 1. Build
cd go-server
go mod tidy
go build -o ghcli-connector-server .
cp ghcli-connector-server ../plugin/servers/go/
# 2. Set up auth — needs the gh CLI itself installed and logged in
# (gh auth login) — see SETUP.md
gh auth status # should succeed before running the server
# 3. Run
./go-server/ghcli-connector-server # serves MCP over stdio
Or make build — see the Makefile for every shortcut
(test, vet, fmt, lint, tidy).
Full walkthrough — including wiring this up as a Claude/Cowork plugin — is in SETUP.md.
🔐 Configuration
Everything is environment variables, passed through by the plugin's
.mcp.json:
| Variable | Purpose | Default |
|---|---|---|
GH_TOKEN / GITHUB_TOKEN |
gh CLI's own token env vars. Leave unset to use whatever gh auth login already configured. |
unset (keychain auth) |
GH_HOST |
Target a GitHub Enterprise hostname instead of github.com. | unset (github.com) |
GHCLI_MCP_ALLOW_WRITE |
"true" to permit state-changing commands at all (still needs confirm=true per call). |
false (read-only) |
GHCLI_MCP_ALLOWED_REPOS |
Comma-separated allowlist of owner/repo values, e.g. "me/proj,me/other". Only enforced when a call passes an explicit -R/--repo flag. |
unset (unrestricted) |
GHCLI_MCP_CLI_PATH |
Path to the gh binary. |
gh resolved via PATH |
🧪 Quality bar
This isn't a toy script — it's got the same checks you'd expect from a production Go service:
- ✅ Unit tests for every input-validation path (
go test ./...) - ✅
go vet+gofmtclean - ✅ golangci-lint (govet, staticcheck, errcheck, gosec, and more)
- ✅ govulncheck — no known vulnerabilities in the dependency graph
- ✅ CodeQL static security analysis on every push
- ✅ End-to-end verified against real GitHub during development — not mocks
- ✅ Dependabot keeps Go modules and Actions current
All of it runs in CI on every push and PR.
🏷️ Releases & versioning
Versions follow semver and are cut automatically by
release-please from
Conventional Commits on main:
fix: ...→ patch (v0.1.0→v0.1.1)feat: ...→ minor (v0.1.1→v0.2.0)feat!: .../BREAKING CHANGE:footer → major (v0.2.0→v1.0.0)
Every merged PR updates a standing "chore(main): release vX.Y.Z" PR with an auto-generated CHANGELOG.md. Merging that PR:
- tags the release and publishes it on GitHub
- builds and attaches zipped, ready-to-install plugin bundles for linux/darwin/windows × amd64/arm64
- regenerates
server.jsonfrom those exact assets (fresh version + SHA-256 hashes) and publishes it to the official MCP Registry viamcp-publisher, authenticated with GitHub OIDC — no stored secrets
See .github/workflows/release-please.yml
and .github/workflows/publish-mcp-registry.yml
(also runnable by hand for an existing tag via workflow_dispatch).
📁 Layout
ghcli-mcp-connector/
├── README.md ← you are here
├── SETUP.md ← step-by-step setup guide
├── CONTRIBUTING.md ← how to contribute
├── CODE_OF_CONDUCT.md
├── SECURITY.md ← vulnerability reporting
├── CODEOWNERS
├── LICENSE ← MIT
├── Makefile ← build / test / lint shortcuts
├── .golangci.yml ← lint rules
├── release-please-config.json ← semver/changelog automation config
├── .release-please-manifest.json
├── server.json ← MCP Registry manifest (regenerated fresh per release by CI)
├── scripts/
│ └── render-server-json.sh ← rebuilds server.json from a release's zip assets
├── .github/
│ ├── workflows/
│ │ ├── ci.yml ← build, vet, test, lint, govulncheck
│ │ ├── codeql.yml ← security scanning
│ │ ├── pr-title.yml ← Conventional Commits PR title check
│ │ ├── release-please.yml ← version PRs, tagging, GitHub releases
│ │ ├── publish-mcp-registry.yml ← publishes server.json to the MCP Registry
│ │ └── rebuild-release-assets.yml ← manual re-attach fallback
│ ├── ISSUE_TEMPLATE/
│ ├── PULL_REQUEST_TEMPLATE.md
│ └── dependabot.yml
├── go-server/ ← the MCP server source
│ ├── main.go
│ ├── main_test.go
│ ├── go.mod / go.sum
│ └── README.md
└── plugin/ ← installable Cowork/Claude plugin
├── .claude-plugin/plugin.json
├── .mcp.json ← holds credentials locally — never commit real ones
├── skills/ghcli-mcp/SKILL.md ← teaches Claude when/how to use the tools
└── servers/go/ ← compiled binary goes here
🤝 Contributing
PRs and issues are very welcome — see CONTRIBUTING.md for the full guide (setup, coding conventions, how to add a new tool) and the Code of Conduct.
main is protected: every change, including the maintainer's, lands via
pull request with CI green. PR titles must follow
Conventional Commits — that's what
drives the automatic versioning above.
Found a security issue? Please follow SECURITY.md instead of opening a public issue.
📄 License
MIT © FerhatDundar
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。