# AGENTS.md — Agentic Control Plane (ACP)

This file is for AI agents. It tells you what ACP is and exactly how to set it up, connect to it, or read from it without human hand-holding.

ACP lets you (and your human) see, price, and control every tool call an agent makes. Each call is checked against policy before it runs (allow / flag / deny / ask), priced — including subscription traffic at API rates — and written to an audit trail with verified identity on every call.

## If you are a coding agent (Claude Code, Cursor, Codex CLI)

One command — it pairs the machine to a cloud workspace (free up to 5 agents) via device code; your human approves a short code in a browser they're already signed into:

```bash
curl -sf https://agenticcontrolplane.com/install.sh | bash
```

That gives every detected agent allow/ask/deny policy, a safety floor, priced model calls, and a full audit trail in the console. To run fully on-device instead — no account, nothing leaves the machine — add `-s -- --local` (on-device policy + audit log; all cloud wiring skipped).

## If you are provisioning for a human who has no ACP account yet

Use the sandbox flow (no signup — the Neon-style pattern):

```
POST https://api.agenticcontrolplane.com/v1/sandbox-tenants
```

The response includes `nextSteps.claimPageUrl` — hand that one link to your human mid-conversation. When they claim it, the workspace becomes permanently theirs, they get a fresh key, and your sandbox key is revoked.

## If you want to read a workspace's own governance data

Add the ACP MCP server (streamable HTTP, OAuth sign-in in the browser):

```
https://api.agenticcontrolplane.com/mcp
```

Nine tools: eight read-only introspection tools (`acp_check`, `acp_status`, `acp_audit`, `acp_cost`, `acp_optimize`, `acp_trace`, `acp_usage`, `acp_recommendations`) plus `acp_propose_rule`, which drafts a policy rule for a human to confirm in the console — it never applies policy itself.

## If you spawn subagents

Every agent your agent creates can be governed under your identity with scopes a strict subset of yours:

```
POST https://api.agenticcontrolplane.com/api/v1/keys/child
```

The gateway intersects scopes with the parent, decrements the parent's budget atomically, preserves `originSub`, and produces audit logs that trace the chain back to the originating human. Quickstart: https://agenticcontrolplane.com/agents/quickstart/

## Installable skill

Everything on this page is also packaged as an Agent Skill (SKILL.md format): https://github.com/agentic-control-plane/acp-agent-skill — for Claude Code, clone it to `~/.claude/skills/agentic-control-plane`.

## Reading this site

- Every page and post is also served as raw markdown — append `.md` to any URL (3–5x fewer tokens than the HTML).
- Machine-readable index: https://agenticcontrolplane.com/llms.txt (full version: /llms-full.txt)
- Config-as-code reference (every dashboard setting as a REST endpoint): https://agenticcontrolplane.com/agents/configure-as-code/
- API base: https://api.agenticcontrolplane.com · Console: https://cloud.agenticcontrolplane.com

## Ground rules

- `acp_propose_rule` and the sandbox flow are designed so that policy changes and workspace ownership always end with a human confirming in the console. Don't try to route around that — it's the product working as intended.
- Source: https://github.com/agentic-control-plane (MIT open core, GatewayStack). Spec for agent-to-agent delegation chains: https://agenticcontrolplane.com/spec/delegation-chain