Install ACP into the coding agent you already run
One command finds the agents on your machine and adds a hook to each. Pick yours below for what the hook sees and what it cannot stop.
Building the agent yourself (CrewAI, LangGraph, Anthropic Agent SDK, OpenAI Agents SDK)? See Frameworks for guided starter code.
The AI tools your team already uses.
Step-by-step install guides for each AI client, with the caveats vendor docs leave out. There are two install paths: a local hook for coding agents, or an MCP server connection for chat clients. What each client lets ACP see and enforce differs, and /coverage lists it.
Using a client we don't list? If it speaks MCP (Windsurf, Continue, ChatGPT’s connectors) or the OpenAI chat-completions API, point it at ACP. Calls that go through the ACP MCP endpoint or proxy are checked and logged. The client's own built-in tools are not.
Want cost too? Hooks cover tool calls. To meter model spend, cache-hit rate, and the loop-vs-leaf cost X-ray, route your model calls through the proxy with one base-URL change.
Your identity provider. Your users.
The installed hooks authenticate with a workspace key. A caller can send a JWT from your IdP instead. ACP checks the signature against your JWKS endpoint, then issuer, audience and expiry, before it trusts any claim. That identity is valid only in your workspace. Your users stay in your directory.
What every integration gets
Whichever client you install, the calls ACP sees go through one policy pipeline and land in one log. Per-client limits are on /coverage.
Keep your clients. Add the checks.
Free to start. Add checks to the AI client you already use in a few minutes.
Already installed? Open your console · For IT teams