内容
<p align="center">
<img src="docs/logo/badge.svg" alt="Steerholm" width="480" />
</p>
<p align="center">
An action plane for your agents, a control plane for you — every action an agent takes is checked against your policy before it reaches a server.<br/>
</p>
<p align="center">
<a href="https://github.com/steerholm/steerholm/releases/latest"><img src="https://img.shields.io/github/v/release/steerholm/steerholm?color=darkgreen&label=version" alt="Version" /></a>
<a href="https://docs.steerholm.ai"><img src="https://img.shields.io/badge/docs-latest-indigo" alt="Docs" /></a>
<a href="LICENSE"><img src="https://img.shields.io/badge/license-MIT-blue" alt="License"></a>
</p>
---
## Install
**Linux / macOS:**
```bash
curl -fsSL https://steerholm.ai/install.sh | bash
```
**Windows (PowerShell):**
```powershell
irm https://steerholm.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](https://github.com/steerholm/steerholm/releases).
## Quick Start
```bash
# 1. Add an MCP server
holm add server git --command "uvx mcp-server-git"
# 2. Add an agent (prints its access key once)
holm add agent my-agent
# 3. Grant it scoped access
holm grant my-agent git --tool "git_log" --args "repo_path=/home/user/projects/**"
```
Then configure your MCP client (Claude Code, VS Code, Cursor, OpenCode) with the
agent's access key as the Bearer token:
```json
{
"mcpServers": {
"steerholm": {
"type": "http",
"url": "http://127.0.0.1:4767/mcp",
"headers": {
"Authorization": "Bearer steer_sk_..."
}
}
}
}
```
## How It Works
```
Agent → HTTP /mcp → Steerholm daemon → MCP Servers
│ │
Bearer auth agent verification
policy enforcement
AUTHORIZATION_DENIED / SERVER_UNAVAILABLE
```
- **Default deny** — no grant means no access
- **Agent from token** — agents cannot self-assert who they are; the access key determines the agent
- **Per-agent policies** — allowlist of servers, tools, and argument constraints
- **Isolation by policy** — one shared daemon; each agent is confined by its policy, not by separate server processes
- **Structured error codes** — `AUTHORIZATION_DENIED` (-31001) and `SERVER_UNAVAILABLE` (-31002)
- **Everything is recorded** — every call, allowed or denied, with the reason: `holm log` (and `holm log --follow` to watch live)
## Documentation
Read the full docs at [docs.steerholm.ai](https://docs.steerholm.ai)
## Contributing
See [CONTRIBUTING.md](CONTRIBUTING.md) for development setup and guidelines.
## Author
[Ismael Kaissy](https://github.com/15m43lk4155y)
## License
This project is licensed under the [MIT License](./LICENSE).
MCP 市场