About

Agentic Control Plane (ACP) is the control layer for AI agents: every tool call an agent makes is checked against your policy before it runs (allow / flag / deny / ask), priced at API rates — including subscription traffic — and written to an audit log. It works regardless of framework, model, or client: one command wires up Claude Code, Codex, or Cursor; SDKs cover CrewAI, LangGraph, OpenAI Agents SDK, Pydantic AI, and a dozen other frameworks.

Who builds it

ACP is built by David Crowe. The open-source core is GatewayStack — six composable npm modules, MIT-licensed — and the delegation-chain primitive is published as an open specification (ADCS). The managed service, ACP Cloud, is free up to 5 initiating agents.

How we work

• Measured, not estimated. Every number we publish comes from something we ran, metered, or scanned ourselves — the canonical index with method and capture date for each is The Data.
• Dogfooded. ACP governs the agents that build ACP. Features ship to our own workspace before they ship to anyone else, and the audit rows in our screenshots are real.
• Built in the open. The modules, the spec, the benchmark (AgentGovBench), and this site are all public repositories.

Contact

Email hello@agenticcontrolplane.com, open an issue on GitHub, or see the contact page for everything else.