Skip to main content

AGP Community Edition

AGP (Agent Governance Plane) is runtime enforcement infrastructure that sits between your AI agents and every MCP tool they can invoke. It is not an SDK, a prompt, or a convention the agent can choose to follow — it is a governance plane the agent cannot route around. Every call passes through an enforcement point your agents can't bypass, and no call reaches a tool backend unless AGP has authenticated the agent, checked its approved operating envelope, evaluated policy, and made an explicit decision to allow it.

Community Edition runs entirely on your machine: single static binaries, state in SQLite under ~/.agp, zero external infrastructure, no telemetry. It is free to use, with no caps on agents, tools, or users.

New to the why?

This guide is the hands-on path. For the architecture and threat model behind it, see AGP Overview and Operational Safety.

The mental model​

Every tool call your agent makes passes through four questions — answered at runtime, on every call:

  1. Who is this agent? — The proxy authenticates a short-lived token issued by Identity. There are no shared credentials and no anonymous agents.
  2. Is this tool in its envelope? — The agent's Behavior Profile determines which tools it can see and call. Tools that are not explicitly granted simply do not exist from the agent's perspective. This is deny-by-default.
  3. What does policy say? — The registry records facts about each tool (read, write, delete, execute). Policy evaluates those facts and decides whether the operation should be allowed, denied, or held for approval.
  4. What is the decision? AGP returns one of three outcomes:

  • allow — execute immediately

  • deny — reject the request

  • hold — pause and wait for human approval

Every decision — allow, deny, or hold — is recorded in an append-only audit trail. Administrative actions such as profile changes, policy updates, tool registrations, and approval decisions are recorded as well.

The result is a complete history of who did what, when, and why.

The path every call takes:

What you'll set up​

By the end of the Get Started guide you will have:

  • The full AGP stack running locally (8 services + console), health-checked.
  • The admin console open in your browser.
  • Your first governed agent, connected to an MCP client (Claude Desktop, Claude Code, Cursor, VS Code, or Codex).
  • A working understanding of why that agent starts out seeing zero tools — and how to grant it the ones it needs.

Prerequisites​

  • macOS, Linux, or Windows (arm64 or amd64).
  • About 5 minutes. No Docker, no database, no cloud account.

The five-command tour​

The whole happy path is five commands. The pages that follow walk through each one in detail.

agp init          # 1. create ~/.agp: secrets, per-service config, CLI profile
agp fetch all     # 2. download the service binaries for your platform
agp start all     # 3. start the stack (health-checked) — console at :27868
agp status        # 4. confirm every service is running and ready
agp setup         # 5. create your first agent + connect a client

Start with Install & first run →