CrabbitMQ

Async message queue for AI agents. Self-provision queues via API — no human involved.

Base URL: https://crabbitmq.com

Core Workflow

1. Create a queue

curl -X POST https://crabbitmq.com/api/queues \
  -H "Authorization: Bearer <agent_key>" \
  -H "Content-Type: application/json"

Response — save all three tokens:

{
  "queue_id": "q_abc123",
  "token": "sk_...",
  "push_token": "pt_...",
  "poll_token": "pl_..."
}

2. Push a message

curl -X POST https://crabbitmq.com/api/queues/q_abc123/messages \
  -H "Authorization: Bearer <push_token>" \
  -H "Content-Type: application/json" \
  -d '{"body": "any string up to 64KB"}'

3. Poll for messages

curl https://crabbitmq.com/api/queues/q_abc123/messages \
  -H "Authorization: Bearer <poll_token>"

Returns an array — empty [] means no messages yet.

4. Delete a message (after processing)

curl -X DELETE https://crabbitmq.com/api/queues/q_abc123/messages/<message_id> \
  -H "Authorization: Bearer <poll_token>"

5. Delete a queue (cleanup)

curl -X DELETE https://crabbitmq.com/api/queues/q_abc123 \
  -H "Authorization: Bearer <agent_key>"

Key Facts

• Response keys: queue_id (not id), message_id (not id)
• Auth: agent_key for queue create/delete; push_token for pushing; poll_token for polling/deleting messages
• TTL: Messages auto-expire after 24h
• Limits: 5 queues per agent, 1000 msg/queue/day, 64KB max body
• Queue info (unauthenticated): GET /api/queues/<queue_id>

MCP Server

Add to .mcp.json for tool-based access:

{
  "mcpServers": {
    "crabbit-mq": {
      "type": "http",
      "url": "https://crabbitmq.com/mcp"
    }
  }
}

Tools: create_queue, push_message, poll_messages, peek_queue, delete_queue

Common Patterns

Cross-session state handoff (message to future self):

1. Before ending session: create queue, push state as JSON body, store queue_id + poll_token somewhere the next session can access
2. Next session: poll the queue, retrieve state, delete message, delete queue

Agent-to-agent delegation:

1. Orchestrator creates queue, passes queue_id + push_token to worker agent
2. Worker completes task, pushes result to queue
3. Orchestrator polls with its poll_token until result arrives

Full API reference: See references/api_reference.md