These five steps take you from a fresh install to an agent that can only reach what
you allow — a server to expose, an agent to authorize, and a scoped grant between
them. Install Steerholm first if you haven’t.
1. Add an MCP server
This registers the git MCP server —
structured git tools (history, diffs, commits) your agent can’t get from a plain
shell. It runs via uvx, which ships with uv. For
remote servers and more, see Add a server.
2. Add an agent
Save the access key — it’s shown only once.
3. Grant access
This lets my-agent read commit history (git_log) — but only for repos under
/home/user/projects/. Every other tool (git_commit, git_reset, …) is denied
by default, so the agent can inspect history but structurally cannot change it.
4. Connect your agent
Point your agent at the endpoint, with the access key as the Bearer token. For
example, Claude Code’s .mcp.json:
Each client stores this differently — see Add an agent for
VS Code, Cursor, and OpenCode. The agent sees only the servers and tools its policy
allows.
Configuring the agent with the key is the current manual step of adding an
agent — Steerholm will set this connection up for you in a future release.
5. Verify it works
Confirm the server is running and see the tools it exposes:
Your agent should now list the tools its policy allows. If none appear,
see Troubleshooting.
What happens at runtime
If the agent tries something unauthorized: