Collision-heavy parallel work
Multiple agents touching the same files produce silent conflicts, duplicated effort, and unnecessary rollbacks.
MACP · Multi-Agent Cognition Protocol · Version 1.0
MACP: the Multi-Agent Cognition Protocol for agents working in parallel.
MACP lets AI agents coordinate inside the execution loop. Each agent announces what it owns, what it changed, and what peers need to know next, all through a shared SQLite-backed bus.
TransportShared SQLite file
DeliveryDurable, poll-based, at-least-once
RoutingChannel broadcast + direct agent delivery
ReferenceTypeScript core + MCP server
Why it exists
Twenty agents on one codebase should not end in merge hell.
Blind execution loops
An agent often discovers critical information too late because no protocol delivers intelligence while it is still working.
Context contamination
A single overloaded agent carries raw logs, retries, and partial reasoning until the useful signal gets buried.
What MACP changes
It turns isolated agents into a coordinated team.
MACP is vendor-agnostic. The schema defines the data model, SQL operations, and prompt-level integration surface. Use the TypeScript MCP server or execute the schema directly from any language with SQLite support.
One shared SQLite file
No broker, no separate control plane, no service to run.
Priority-ranked delivery
Messages arrive as `info`, `advisory`, `steering`, or `interrupt` based on urgency.
Own files before editing
Agents register, claim work, and expose progress before they touch the code.
Durable, poll-based semantics
At-least-once delivery with ACK state and byte-budget pruning built into the protocol.
01
Observers and specialists maintain clean, purpose-built contexts.
02
Findings are routed through one durable bus with ranked urgency.
03
The executing agent polls mid-task and adapts before mistakes compound.
Inside the loop
Situational awareness arrives while work is still in motion.
- Register each agent against a shared database path.
- Join a channel for the current project or execution thread.
- Broadcast file ownership, findings, and test status as work progresses.
- Poll during execution so every agent can adjust in real time.
Where it applies
Any environment where data moves fast and decisions cannot wait.
Multi-agent software deliveryFraud and anomaly monitoringDistributed systems operationsResearch synthesis and analysisSupply chain and logistics routingTrading and market response systems
Quickstart
Deploy the reference server and point agents at the same DB.
Build and run
npm install
npm run build
MACP_DB_PATH=/tmp/macp_team.db \
MACP_AGENT_ID=agent-alpha \
MACP_AGENT_NAME=Alpha \
MACP_DEFAULT_CHANNEL=case-001 \
node build/src/server.jsMCP configuration
{
"mcpServers": {
"macp": {
"command": "node",
"args": ["path/to/macp/build/src/server.js"],
"env": {
"MACP_DB_PATH": "/tmp/project.macp.db",
"MACP_AGENT_ID": "agent-1",
"MACP_AGENT_NAME": "Frontend Specialist",
"MACP_DEFAULT_CHANNEL": "project-x"
}
}
}
}Contact
Talk to us about deploying MACP in real systems.
Reach out if you want to evaluate MACP in a multi-agent workflow, compare integration approaches, or discuss coordination problems in production environments.
Good reasons to reach out
- Adapting MACP to an existing agent stack
- Evaluating coordination and file-ownership semantics
- Discussing protocol fit for production systems
Send a message
We will reply by email.
Use the form to start a conversation about MACP deployment, evaluation, or integration.
MACP
Coordination is now part of the protocol surface.
If multiple agents can work your problem in parallel, they also need a shared cognitive layer. MACP provides it with one file and no infrastructure overhead.