🚀 Shuttle
Secure SSH gateway for AI assistants
接入你的智能体
從發布者安裝。不是 PlugMCP 端點。
遠端 URL 還是本機指令? 兩種安裝方式的差異
Shuttle lets AI assistants (Claude Code, Cursor, etc.) securely execute commands on your remote SSH servers — with connection pooling, session isolation, command safety rules, and a web audit panel.
Getting Started · MCP Tools · Web Panel · Security Rules · Docs · 中文文档
When AI coding assistants need to operate remote servers (run tests on GPU machines, deploy to staging, check logs), they need a secure bridge. Shuttle provides:
# Recommended: install CLI once (tools bin on PATH)
uv tool install shuttle-mcp
shuttle --help
# Or run without installing (stdio / one-off)
uvx shuttle-mcp --help
# Older PyPI wheels without the `shuttle-mcp` script:
# uvx --from shuttle-mcp shuttle --help
shuttle node add
# Follow the prompts: name, host, username, password/key
Claude Code / Cursor (stdio mode):
// .mcp.json
{
"mcpServers": {
"shuttle": {
"command": "uvx",
"args": ["shuttle-mcp"]
}
}
}
Service mode (with Web UI):
# Start the service
shuttle serve
# Then configure your AI client with the URL
// .mcp.json
{
"mcpServers": {
"shuttle": {
"url": "http://localhost:9876/mcp/"
}
}
}
That's it. Your AI assistant can now execute commands on your remote servers.
Both modes share the same SQLite database — commands logged in CLI mode are visible in the Web UI when you switch to service mode.
AI assistants get these tools automatically:
You: Check the GPU usage on my training server
AI: → ssh_run(node="gpu-server", command="nvidia-smi")
AI: Your GPU server has 7x A100-80GB, all idle at 0% utilization.
You: Start a training run
AI: → ssh_run(node="gpu-server", command="cd /workspace && python train.py")
AI: Training started. Epoch 1/10... (working directory preserved automatically)
Commands are evaluated against a 4-level security system:
Default rules are seeded on first startup. Customize via Web UI or directly in the database.
Different servers can have different rules:
Global: sudo .* → confirm
GPU Server: sudo .* → allow (trusted environment)
Prod Server: DROP TABLE → block (extra protection)
Start with shuttle serve, open http://localhost:9876:
The Web UI requires a bearer token (displayed when you run shuttle serve).
# MCP Server
shuttle # Start MCP server (stdio mode)
shuttle serve # Start service mode (MCP + Web)
shuttle serve --port 8080 # Custom port
shuttle serve --host 0.0.0.0 # Bind to all interfaces
# Node Management
shuttle node add # Add node interactively
shuttle node list # List all nodes
shuttle node test <name> # Test SSH connection
shuttle node edit <name> # Edit a node
shuttle node remove <name> # Remove a node
# Configuration
shuttle config show # Display current config
All settings can be overridden with environment variables (prefix SHUTTLE_):
SHUTTLE_DB_URL=postgresql+asyncpg://user:pass@host:5432/shuttle shuttle serve
Requires: uv pip install asyncpg (install into the same environment that runs Shuttle)
# Clone and install
git clone https://github.com/enwaiax/shuttle.git
cd shuttle
uv sync
# Run tests
uv run pytest tests/ -v
# Lint
uv run ruff check src/ tests/
# Frontend dev (hot reload)
cd web && npm install && npm run dev
# Backend: uv run shuttle serve (in another terminal)
Developer ↔ AI Assistant ↔ Shuttle (MCP) ↔ SSH ↔ Remote Servers
│
┌─────────┴──────────┐
│ Core Engine │
│ ├ ConnectionPool │
│ ├ SessionManager │
│ ├ CommandGuard │
│ └ SQLAlchemy ORM │
└────────────────────┘
Service mode: Single ASGI app serving both MCP (at /mcp/) and Web UI (at /) on the same port.
Secure SSH gateway for AI assistants — MCP server with connection pooling, session isolation, command safety, and web audit panel
工具
尚無工具清單。遠端伺服器會在首次健康檢查後取得工具清單。
常見問題
在本頁複製適用於 Cursor、Claude Desktop、Claude Code 或 VS Code 的 stdio 設定。它使用發布者的安裝方式,而不是 PlugMCP 代理伺服器。
這筆記錄未列出必要的環境變數。
Shuttle 透過 stdio 在本機執行。
上次檢查時間:2026-10-28。 Secure SSH gateway for AI assistants — MCP server with connection pooling, session isolation, command safety, and web audit panel