Agentic Control Plane: facts
Last updated 2026-11-17. Each line restates a page linked from it.
What it is
Agentic Control Plane (ACP) is a hook for AI coding agents that checks each tool call the agent’s hook can see before it runs, and records it in one log (what each agent exposes). By default the destructive shell calls wait for a person to approve them, and everything else runs.
The name
“ACP” on this site means Agentic Control Plane. Two unrelated projects use the same acronym:
- Agent Client Protocol: the JSON-RPC standard Zed introduced for connecting editors and clients to coding agents. DeepSeek Harness runs it as
dsh --profile acp; the community package isdeepseek-harness-acp. - Agent Communication Protocol: IBM Research / BeeAI, for agent-to-agent messaging.
ACP vs ACP routes between them.
Coding agents it covers
The installer sets up 14 coding agents: Claude Code, Codex, Cursor, GitHub Copilot, OpenCode, OpenClaw, Hermes, DeepSeek Harness, Qwen Code, Antigravity, Grok Build, Pi, Prime Agent, Muse Code.
fx installs from npm (@agenticcontrolplane/fx) as middleware on fx acp, not through the installer. Agent frameworks use a wrapper instead of a hook: CrewAI, LangGraph, Deep Agents, OpenAI Agents SDK, Anthropic SDK, Google ADK.
What the hook sees on each agent, and its limits there, is in the coverage matrix.
What it does by default
- By default the destructive ones (a force-push, destructive SQL through a database client,
curl | sh, a recursive delete outside the project, and a few more) are blocked until you approve.rm -rfof your home or root,mkfs,ddonto a disk, fork bombs and a few more are always refused. (What the installer writes) - The holds that wait for a person read shell commands only, in Bash and PowerShell. A few refusals also cover file edits: an agent cannot edit ACP’s own config, its hook settings or your shell rc files. Other file edits and MCP tools are recorded and follow the rules you write by tool name.
- On Claude Code, when the person at the keyboard is the workspace owner or an admin, a safety-floor hold is a prompt in the session; a member’s floor hold goes to an admin on the console. Holds from your own rules, and any hold in a headless run or with prompts bypassed, are blocked and approved on the console or from the email. On the other harnesses the call is blocked, you approve it on the console or from the email, then ask the agent to retry. Grok Build in default and plan modes shows its own prompt instead.
- On DeepSeek Harness a deny is reported to the agent but not enforced by the plugin yet.
- Your own rules are allow, ask first or never, by tool, and apply to every agent. A new connected workspace starts in audit mode: calls are recorded and your rules do not block until you turn enforcement on. The safety floor holds and refuses from the first call either way, unless it was switched off at install or on the Policies page.
What a connected hook sends
- Connected, the hook sends ACP each call’s tool name and arguments, and up to 200 KB of its output. On Claude Code the hook also reads the local session transcript to price model turns; it sends token counts, never message text.
- By default the stored record is metadata: tool, who ran it, the decision and reason, latency, tokens and cost, plus the harness and its version, the working directory, session and call ids, and the agent name your harness reports (Trust). Command text and output are not stored unless a workspace admin turns on short previews. A held call keeps up to 500 characters of its arguments for the approver. A credential in a known format is redacted before the record is written.
- Add
--localand nothing is sent: the rules and the log stay on this machine, with no account. Rules live in~/.acp/policy.jsonand the log in~/.acp/audit.jsonl.--localsets up Claude Code, Codex, Cursor, GitHub Copilot, Qwen Code; the other harnesses need a free workspace. On OpenClaw,--localinstalls the plugin, but it checks nothing without a workspace key.
Price
Free for you. $100 a month for a team.
- Free: $0 (5 agents · unlimited calls)
- Team: $100/mo (25 agents · unlimited calls)
- Scale: $1,000/mo: More than 25 agents? Scale covers 250 for $1,000 a month, with everything in Team.
- Enterprise: Let’s talk: Need SSO, self-hosting, or a DPA? Those are available on request, with no cap on agents.
Details: Pricing.
Install
# macOS, Linux, WSL
curl -fsSL https://agenticcontrolplane.com/install.sh | bash
# Windows (PowerShell)
irm https://agenticcontrolplane.com/install.ps1 | iex
# no account, nothing sent
curl -fsSL https://agenticcontrolplane.com/install.sh | bash -s -- --local
The installer lists the coding agents it found and asks which to cover before it writes anything. acp-uninstall removes it; after a --local install, run curl -fsSL https://agenticcontrolplane.com/uninstall.sh | bash. Every file it writes: What the installer writes.
Source
- Code: github.com/agentic-control-plane. The hook (
govern.mjs) comes from claude-code-acp-plugin. - The installer is plain text you can read before running it: /install.sh.
Running coding agents? ACP records each tool call its hook sees and makes the dangerous shell commands (a force-push, destructive SQL, curl | sh) wait for you. One command, free up to 5 agents:
curl -fsSL https://agenticcontrolplane.com/install.sh | bash
irm https://agenticcontrolplane.com/install.ps1 | iex