使用说明
[!IMPORTANT] MCP Harbour is now Steerholm. This repository is archived and no longer maintained — development continues at steerholm/steerholm · steerholm.ai. Existing MCP Harbour installs stay on the final
v0.1.3release here and will not auto-update.
Dock your servers once, control which agents can access which tools, and manage everything from a single place.
Built as an implementation of the GPARS plane boundary.
Install
Linux / macOS:
curl -fsSL https://mcpharbour.ai/install.sh | bash
Windows (PowerShell):
irm https://mcpharbour.ai/install.ps1 | iex
This downloads the binary, registers a per-user background daemon that runs as you, and starts it. No admin rights, no Python or package manager required.
Or download binaries directly from GitHub Releases.
Quick Start
# 1. Dock an MCP server
harbour dock --name filesystem \
--command "npx -y @modelcontextprotocol/server-filesystem /home/user/projects"
# 2. Create an identity
harbour identity create my-agent
# 3. Grant permissions
harbour permit allow my-agent filesystem --tool "*" --args "path=/home/user/projects/**"
Then configure your MCP client (Claude Code, VS Code, Cursor, OpenCode):
{
"mcpServers": {
"harbour": {
"url": "http://127.0.0.1:4767/mcp",
"headers": {
"Authorization": "Bearer harbour_sk_..."
}
}
}
}
How It Works
Agent → Streamable HTTP /mcp → Harbour Daemon → MCP Servers
│ │
Bearer auth identity verification
policy enforcement
AUTHORIZATION_DENIED / SERVER_UNAVAILABLE
- Default deny — no policy means no access
- Identity from token — agents cannot self-assert their identity
- Per-agent policies — whitelist of servers, tools, and argument constraints
- Isolation by policy — one shared daemon; each agent is confined by its policy, not by separate server processes
- GPARS error codes —
AUTHORIZATION_DENIED(-31001) andSERVER_UNAVAILABLE(-31002)
Documentation
Read the full docs at docs.mcpharbour.ai
Contributing
See CONTRIBUTING.md for development setup and guidelines.
Author
License
This project is licensed under the MIT License.
