Skip to main content
Steerholm is one surface with two faces. To your agents it’s an action plane — the single place they connect to and act through. To you it’s a control plane — where you register the agents and the MCP servers they might use, then grant each agent policy-scoped access to exactly the capabilities it needs. Nothing reaches a server that you didn’t allow; you stay in control of what every agent can touch.

​
The three elements

Everything on the control plane is one of three things:

Agents

The AI tools you govern (Claude Code, Cursor, VS Code, OpenCode). Each has a name, an access key, and a policy.

Servers

The MCP servers you add. Each one exposes tools — the capabilities an agent might use.

Grants

Controlled access from an agent to a server’s capabilities, scoped by policy. No grant means no access.
Two kinds of thing you add (agents, servers) and one governed edge you draw between them (a grant). That’s the whole model.

​
How you operate it

Every command reads as manage an agent or control its access:
Grants are additive and default-deny: an agent starts with zero access and only ever gets what you grant. See Permissions for the tool and argument syntax.

​
A grant is a governed connection

When you grant an agent access to a server, you’re wiring a connection routed through Steerholm. The agent never talks to the server directly — it connects to Steerholm, and Steerholm forwards each call to the server only where the grant and its policy allow. The grant is the edge; Steerholm is what the traffic flows through and gets checked at. (See Architecture for how that boundary is enforced.)

​
Bringing an agent online

Registering an agent and pointing it at Steerholm are two steps today:
  1. holm add agent <name> registers the agent and issues its access key.
  2. You configure the agent with that key so it opens its connection to Steerholm — see Add an agent.
  3. holm grant ... gives the agent access to the servers it should reach.
Step 2 is the current manual last mile. It’s part of adding an agent — just not yet automated. We’re working toward Steerholm setting up that connection for you (and launching agent sessions directly), so the copy-the-key step goes away while the model above stays exactly the same.