CodeSherpa
Remote MCP server enabling AI clients to securely inspect repositories over SSH with read-only tools.
README
<p align="center"> <img src="https://raw.githubusercontent.com/boridon/code-sherpa/main/docs/logo.svg" width="220" alt="CodeSherpa logo"> </p>
<h1 align="center">CodeSherpa</h1>
<p align="center">AI-guided repository exploration over SSH via MCP</p>
What Is CodeSherpa?
CodeSherpa is a remote MCP server that lets AI clients inspect repositories over SSH with a strict read-only toolset.
Supported MCP tools:
healthcheck_remotelist_filesread_filesearch_codegit_statusgit_diffgit_log
Compatible MCP clients:
- ChatGPT (custom connectors)
- Claude Desktop
- Cursor
- Other MCP-compatible agents
Quick Start
Clone and start CodeSherpa locally.
git clone https://github.com/boridon/code-sherpa.git
cd code-sherpa
cp .env.example .env
docker compose up -d
Then connect your MCP client to:
https://your-domain.example/mcp
Architecture
MCP Client
|
v
CodeSherpa (HTTPS)
|
v
SSH (read-only user)
|
v
Private repository host
Key points:
- Repository data stays on the SSH target host.
- CodeSherpa exposes only read-only MCP tools.
- Path traversal and sensitive path segments are blocked.
- OAuth access tokens and legacy fixed bearer tokens are supported.
Security Model
- Use a read-only SSH user (no sudo).
- Denied path segments include
.git,.env,node_modules, and similar sensitive paths. - Absolute paths and
..traversal are rejected. /mcprequiresmcp:readscope for OAuth access tokens.- Legacy fixed bearer token auth can remain enabled for internal testing.
Note: OAuth sessions, authorization codes, and tokens are in-memory in the current implementation. They are reset when the container restarts.
Docker Deployment
1. Clone and prepare
git clone https://github.com/boridon/code-sherpa.git
cd code-sherpa
cp .env.example .env
mkdir -p secrets
2. Add SSH secrets
- Put your private key at
secrets/id_ed25519 - Generate known hosts:
ssh-keyscan -H <ssh-host> > secrets/known_hosts
chmod 600 secrets/id_ed25519
chmod 644 secrets/known_hosts
3. Start the service
docker compose build
docker compose up -d
docker compose ps
curl http://127.0.0.1:8787/health
Cloudflare Tunnel
You can run CodeSherpa with Cloudflare in two ways.
A. Sidecar container (token mode)
docker compose -f docker-compose.yml -f docker-compose.cloudflare.yml up -d
This mode uses CLOUDFLARE_TUNNEL_TOKEN from .env.
B. Config file mode (cloudflared-config.yml)
Example ingress:
ingress:
- hostname: code-sherpa.example.com
service: http://localhost:8787
- service: http_status:404
If cloudflared is not installed on your host:
- Debian/Ubuntu:
sudo apt-get install cloudflared - RHEL/CentOS/Fedora:
sudo dnf install cloudflared - macOS (Homebrew):
brew install cloudflared
OAuth for MCP Connectors
CodeSherpa includes a minimal built-in OAuth authorization server for connector setup flows.
OAuth discovery endpoints:
GET /.well-known/oauth-authorization-serverGET /.well-known/openid-configuration
OAuth endpoints:
GET /authorizePOST /tokenGET /loginPOST /loginGET /oauth/consentPOST /oauth/consent
OAuth profile:
- Grant type: Authorization Code + PKCE (
S256) - Scope:
mcp:read - Public client support: yes (
token_endpoint_auth_method=noneallowed) - Refresh token: supported
ChatGPT Connector Values (Example)
Use example values like these:
- MCP endpoint:
https://code-sherpa.example.com/mcp - Issuer:
https://code-sherpa.example.com - Authorization endpoint:
https://code-sherpa.example.com/authorize - Token endpoint:
https://code-sherpa.example.com/token - Scope:
mcp:read
Environment Variables
Use .env.example as the baseline.
Required:
SSH_HOSTSSH_PORTSSH_USERNAMEREPO_ROOTMCP_BEARER_TOKEN(for optional legacy/manual testing)OAUTH_ISSUER_BASE_URLOAUTH_LOGIN_USERNAMEOAUTH_LOGIN_PASSWORDOAUTH_SESSION_SECRET
Optional/common:
PORT(default8787)MCP_SERVER_NAME(defaultcode-sherpa)MCP_SERVER_VERSION(default0.1.0)OAUTH_COOKIE_SECURE(defaulttrue)MAX_FILE_BYTES,MAX_SEARCH_RESULTS,MAX_LOG_COMMITS,MAX_RESPONSE_CHARS
Example .env snippet (safe placeholders):
PORT=8787
MCP_SERVER_NAME=code-sherpa
SSH_HOST=ssh-host.example.internal
SSH_PORT=22
SSH_USERNAME=repo_reader
REPO_ROOT=/srv/repos/project
OAUTH_ISSUER_BASE_URL=https://code-sherpa.example.com
OAUTH_LOGIN_USERNAME=replace-me
OAUTH_LOGIN_PASSWORD=replace-me
OAUTH_SESSION_SECRET=replace-with-long-random-secret
MCP_BEARER_TOKEN=replace-with-long-random-token
Minimal Verification
1. OAuth discovery
curl -i http://127.0.0.1:8787/.well-known/oauth-authorization-server
curl -i http://127.0.0.1:8787/.well-known/openid-configuration
2. Legacy bearer test
curl -i -X POST http://127.0.0.1:8787/mcp \
-H "Authorization: Bearer ${MCP_BEARER_TOKEN}" \
-H 'Content-Type: application/json' \
-H 'Accept: application/json, text/event-stream' \
-d '{"jsonrpc":"2.0","id":"1","method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"curl","version":"0.0.1"}}}'
3. OAuth token exchange
After browser login + consent, exchange the authorization code:
curl -i -X POST http://127.0.0.1:8787/token \
-H 'Content-Type: application/x-www-form-urlencoded' \
--data-urlencode 'grant_type=authorization_code' \
--data-urlencode 'code=<authorization-code>' \
--data-urlencode 'redirect_uri=<same-redirect-uri-used-at-authorize>' \
--data-urlencode 'code_verifier=<pkce-code-verifier>' \
--data-urlencode 'client_id=<client-id>'
Project Structure
code-sherpa
├── src/
│ ├── index.ts
│ ├── oauth.ts
│ ├── pkce.ts
│ ├── session.ts
│ └── token-store.ts
├── docs/
│ └── logo.svg
├── Dockerfile
├── docker-compose.yml
├── docker-compose.cloudflare.yml
├── cloudflared-config.example.yml
├── cloudflared-config.yml
├── .env.example
├── .gitignore
├── LICENSE
└── README.md
License
MIT
Contributing
Issues and pull requests are welcome.
If CodeSherpa is useful to you, consider giving the repository a star on GitHub.
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。