Skip to the content

Quick start

Install a policy, run an agent through AgentGarm and see what it did.

Not installed yet? Install AgentGarm first: every step here assumes a running install.

1. Write a policy

Install the default pack for your agent, in audit mode: it decides and records, and blocks nothing yet. A root install reads the system's policy directory; a no-root install reads yours.

sudo agentgarm policy init claude-code --dir "/Library/Application Support/AgentGarm/agents.d"
sudo agentgarm policy init claude-code --dir /etc/agentgarm/agents.d
agentgarm policy init claude-code

agentgarmd reads its policy directory when it starts, so restart it after adding a pack: how.

2. Run the agent through AgentGarm

agentgarm run starts the agent in the sandbox, with its model traffic through the local proxy and every event tied to one task:

agentgarm run -- claude

Start it from a plain terminal: a shell that Claude Code or the Claude desktop app started exports ANTHROPIC_BASE_URL, which routes around the proxy and rates the task C. From such a shell, run env -u ANTHROPIC_BASE_URL agentgarm run -- claude.

Sign in with an API key: export ANTHROPIC_API_KEY before you launch, and answer Yes when Claude Code asks whether to use it. A claude.ai subscription login is never routed, so with one Claude Code reports Not logged in (more). On a first run the launcher lists your existing hooks, plugins and MCP servers as awaiting approval and launches anyway; review them with agentgarm approve.

3. See what it did

agentgarm status
agentgarm ui --open

The dashboard shows each task's decisions, approvals, alerts and cost.

4. Enforce

Once the audit shows the decisions you expect, check the policy against your own history, then switch it from mode: audit to mode: enforce. Policies.

Where next