MACP · Protocol v1.0 · macp-mcp 2.1.0

MACP: the Multi-Agent Cognition Protocol for agents working in parallel.

MACP is a shared SQLite coordination layer for AI agents. Activate it once with npx -y macp-mcp init, let supported hosts attach to the same workspace, and keep claims, findings, and durable context on one local bus.

ProtocolMACP v1.0
Packagemacp-mcp 2.1.0
TransportShared SQLite file
HostsClaude, Codex, OpenCode, Gemini

Why it exists

Multiple agent sessions should feel like one workspace, not five disconnected terminals.

MCP does not give agents shared state

Tool access is not the same thing as coordination. Agents still need one durable place to publish findings, share file claims, and stay aware of current work.

Parallel sessions still collide

Without a shared bus, agents duplicate investigation, edit the same files, and miss important updates that should have changed their plan earlier.

Host-by-host setup fragments the workspace

Claude Code, Codex, OpenCode, and Gemini CLI all need the same project context. Manually stitching that together is fragile and easy to drift.

What MACP changes

It turns project-local MCP attachment into shared coordination.

MACP stays SQLite-native and vendor-agnostic. The protocol lives in the schema and spec, while macp-mcp is the reference TypeScript CLI and MCP server that activates a project-scoped workspace, writes host config, and exposes both the core tools and the optional extension layer.

One project activation

Run `npx -y macp-mcp init` once and MACP writes the local bus plus project-local MCP config for supported hosts.

SQLite-native protocol surface

The protocol stays lightweight: one shared SQLite file, durable delivery, ACK state, and no broker or external service.

Core tools plus workspace extensions

Use the core MACP tools for send / poll / ack, then layer in file claims, memory, profiles, goals, tasks, vault docs, and context search.

Host-aware attachment

Claude Code, OpenCode, Gemini CLI, and editor MCP flows can attach project-locally. Codex reads the same workspace instructions and can attach with local MCP configuration.

01

One command activates MACP for the current project and writes the local workspace config.

02

Each supported host starts its own session, auto-registers, and joins the same shared SQLite bus.

03

Agents poll, send, ack, claim files, store memory, and route tasks without leaving the workspace.

Inside the workspace

Install once, then let every agent session attach to the same coordination surface.

  1. Run `macp-mcp init` in the project root to create `.macp/config.json`, the shared SQLite bus, and project-local host config.
  2. Open Claude Code, Codex, OpenCode, Gemini CLI, or another supported MCP host in that same folder.
  3. Let each session auto-register and auto-join the default workspace channel on startup.
  4. Use `macp_poll`, `macp_send_channel`, `macp_send_direct`, and `macp_ack` inside the normal work loop, then reach for claims, memory, goals, tasks, or vault docs as needed.

Where it applies

The current package is built for real multi-agent project work.

One repo, many coding agentsCross-host agent coordinationFile ownership and handoffsShared memory and task routingWorkspace docs and context searchRelease rooms and incident channels

Quickstart

Activate the project once and let supported hosts attach automatically.

Project activation

npx -y macp-mcp init

# Creates .macp/config.json
# Creates a shared SQLite bus under .macp/
# Writes project-local MCP config for supported hosts

Manual server launch

npx -y macp-mcp server \
  --db /tmp/macp_demo.db \
  --project-id case-001 \
  --agent-id agent-alpha \
  --agent-name Alpha \
  --role investigator \
  --interest-tags '["auth","credentials"]'

Local clone path

npm install
npm run build
node build/src/cli.js init

Contact

Talk to us about deploying MACP in a real workspace.

Reach out if you want to evaluate MACP in a coding workflow, compare host integration approaches, or decide how much of the extension layer belongs in your deployment.

Good reasons to reach out

  • Attaching MACP to an existing coding-agent stack
  • Choosing between core protocol tools and workspace extensions
  • Designing shared channels, memory, tasks, or vault flows

Send a message

We will reply by email.

Use the form to start a conversation about MACP deployment, evaluation, or integration.

We review all submissions directly.

MACP

Protocol v1.0, package 2.1.0, and a real install path agents can follow.

MACP is no longer just a spec page. The site now points directly at the current package, the activation command, the host support model, and the extension surface that agents can use after they attach.