ACP Agent Control Plane

Free and open source · MIT license

All your coding agents.
One memory.

Bring Claude Code, Codex, free and local models. They plan together, work in parallel and hand off without losing the thread. Nothing in your project changes until you approve it.

In development · the desktop app is not released yet

How it works

ACP doesn't replace your agents. It connects them, gives them one shared memory, and keeps a person in charge of what changes.

  1. 01

    Connect what you have

    Sign in to Claude Code, Codex or Antigravity in their own apps. ACP starts them for you and never sees your password or tokens. No paid plan? Use free models through OpenCode.

  2. 02

    Pick an orchestrator

    Star one or two agents. They talk through your goal, split it into tasks and hand each task to the model best suited to it.

  3. 03

    Work runs in parallel

    Each task runs in its own Git worktree. If an agent stops or its plan runs low, another picks up from the latest checkpoint, not the whole chat.

  4. 04

    You approve every change

    Each change is tested and scanned and arrives with a report. Approved changes merge one at a time, with checks run again after each.

shared memory

Switch agents without writing a briefing.

Claude, ChatGPT and Gemini each keep their own memory, locked in their own app. ACP keeps one memory for every agent you connect, owned by you.

Project memory lives in your repository, so it's reviewed like code and works even without ACP running. ACP also writes it into the AGENTS.md and CLAUDE.md files your agents already read.

LayerWhat it holds
YouPreferences for every project
ProjectGoals, rules and decisions in .acp/knowledge.md
Folder rules"In src/api, always…", matched by path
TaskCheckpoints and handoff.md

The narrower layer wins, so a folder rule beats a project rule.

review and safety

Evidence first, opinions labelled.

Every change runs the same checks, whichever model wrote it. Tool results are shown as evidence. Comments from AI reviewers are marked as opinion, so you know which is which.

  • ✓ Tests and types
  • ✓ Stays inside its scope
  • ✓ No secrets committed
  • ✓ Known-vulnerable packages
  • ✓ Security patterns
  • ✓ Dead code and duplicates

Example patch report

Problem

New accounts stored the password as plain text.

Change

Hash the password before saving, using the existing helper.

Scope

src/api/account/** · no new dependencies

Verification

npm test · 41 passed
new test fails on main, passes here

Use your plan, or start free.

Paid plans run in the maker's own app, signed in by you. Free models run through OpenCode, so they can edit files and run commands too.

Claude CodeCodexAntigravityOpenCode NVIDIA · free tierOllama · local

Where ACP is today.

We'd rather tell you plainly than overpromise.

Works now

  • Local dashboard for tasks, checkpoints and patches
  • Patch review: validate, approve, apply
  • Each task in its own Git worktree
  • Runs in Docker on your computer

Being built

  • Connecting Claude Code, Codex, Antigravity and OpenCode
  • Shared memory in layers
  • Orchestrators and handoffs

Planned

  • Desktop app for macOS, Windows and Linux
  • Screen recordings as evidence in reports
  • Design library as skills
# Try it with Docker git clone https://github.com/henryuks129/agent-control-plane.git cd agent-control-plane docker compose up -d # then open http://127.0.0.1:4317

Bring the agents you already pay for.