An initiating agent is an identity that starts work — your coding agent, a deploy bot, a scheduled job. The subagents and delegation chains under it are covered, unlimited. No call meters, no seats, no tax on coverage.
Think hosts and containers: you pay for the host, containers are free. Count the agents you'd name if someone asked "what agents do you run?" — that's your bill. Every workspace's console shows the named list, so the number is one you can verify yourself.
Per-call pricing taxes coverage — it punishes you for routing your whole fleet through the control plane, which is the one behavior everyone should want. Per-seat pricing charges by the humans, and the point of autonomous agents is that fewer humans are in the loop. The fleet is what actually grows, so the fleet is what we price — in wide bands, so adding one agent never changes your bill.
The enforcement modules — identity, PII redaction, policy, limits, routing, audit — are MIT-licensed npm packages you can read and run yourself, and the on-device mode wires them into your agent harness with one command, no account, nothing phoning home:
curl -sf https://agenticcontrolplane.com/install.sh | bash -s -- --localWindows — in PowerShell (connected install; --local is POSIX-only today): irm https://agenticcontrolplane.com/install.ps1 | iex
Local mode uses the same policy format and audit schema as ACP Cloud — start on your laptop, attach a workspace later without reinstalling. The hosted control plane — console, team approvals, agent-to-agent control, cost X-ray, multi-tenant isolation — is how you run those layers in production. It is not open source; a self-hosted or VPC deployment of the full gateway is available on Enterprise, on request.
The six layers, if you want them individually: identifiabl (identity), transformabl (PII), validatabl (policy), limitabl (limits & runaway detection), proxyabl (identity-aware routing), explicabl (audit).
A durable identity that starts work — a person's coding agent (Claude Code, Cursor, Codex, any client), a named background agent, a CI bot, a scheduled job. It's counted once per month no matter how many sessions it opens or how many machines it runs on. Reinstalls and key rotations don't fork an agent's identity, and our own install-verification probe never counts.
Free, always, on every plan — and deliberately so. ACP enforces policy down the entire delegation chain; charging for every agent in the chain would punish exactly the architectures we want you to build. Fan out as wide as you like: only the agent at the root of the chain counts.
No call caps on any plan. The only number that matters is how many agents start work in a month.
Nothing breaks — blocking your agents mid-task is the one thing a control plane must never do to you. Agents beyond your band keep working, but they run ungoverned: no policy, no audit trail, no cost tracking, with a loud notice on every call so nobody finds out later. Upgrading puts them back under policy and audit immediately.
No. Your model spend stays on your own keys and subscriptions, always — ACP never touches or marks up your tokens. We meter and X-ray what your agents cost you; we don't take a cut of it.
Self-hosted uses the open-source npm modules in your own infrastructure. Managed (ACP Cloud) adds the web console, built-in integrations, agent-to-agent control, multi-tenant isolation, and we handle deployment and scaling.
Yes. Free needs no card. Paid plans are flat monthly — upgrade, downgrade, or cancel as your fleet changes.
ACP is not certified for SOC 2, HIPAA, or ISO 27001, and we say so plainly on our trust page. What ACP provides today is the evidence layer those frameworks ask for: identity-attributed logs of every tool call, redaction findings, policy decisions, and per-user attribution — audit exports built to serve as evidence in your own certification work.
Start free with up to 5 agents. Grow into a band when the fleet does — subagents free, calls unlimited, always.