Last year I connected an app to an LLM with a MCP connector. Turned out that authenticating the LLM user in the app backend was surprisingly hard.
That was the canary in the coalmine. If it's hard to authenticate actual users:
- what about their agents? - what about downstream governance? Permissions, limits, audit logs
ACP is a governance layer that sits in front of AI coding agents like Claude Code and OpenClaw. It runs on every tool call (Bash, Read, Write, file edits, web fetches, MCP, API calls). Every call is logged and optionally policy checked before execution.
It works by hooking into your agent's tool pipeline. For Claude Code, it's a PreToolUse hook (~200ms). For OpenClaw, it's a before_tool_call plugin at priority 0. The plugins are MIT-licensed and all governance logic runs server-side.
You get an audit trail and can set permissions and limits. Allow/ deny by tool or type of agent.
One-command install for Claude Code and OpenClaw:
curl -sf https://agenticcontrolplane.com/install.sh | bash