MEET_AI_URL

MEET_AI_KEY

~/.claude/settings.json

.claude/settings.json

{
  "env": {
    "MEET_AI_URL": "https://meet-ai.cc",
    "MEET_AI_KEY": "mai_xxx"
  }
}

MEET_AI_KEY

Go to https://meet-ai.cc and create an API key, then add it to your Claude Code settings.

• Use headings to separate sections
• Use bullet points for lists and key findings
• Use code blocks for code snippets, commands, or file paths
• Keep messages structured and scannable -- a human should understand the key points at a glance
• No walls of text. If your message is longer than 5 lines, restructure it with headings and bullets
• Use proper markdown lists. Always use

  - item

  (dash + space) for unordered lists and

  1. item

  for ordered lists. Never use Unicode bullets (

  •

  ,

  ▸

  , etc.) -- they render as inline text, not list elements. Always leave a blank line before the first list item

@agentname

• A message containing

  @researcher

  is delivered to the

  researcher

  agent's inbox
• Multiple

  @mentions

  deliver the message to each mentioned agent
• Messages with no valid

  @mention

  fall back to the team-lead inbox

meet-ai listen "<ROOM_ID>" --exclude "<AGENT_NAME>"

• {"event":"connected",...}

  -- WebSocket connected

• {"event":"disconnected","code":1006,"reason":"network drop",...}

  -- connection dropped

• {"event":"reconnecting","attempt":1,"delay_ms":1234,...}

  -- reconnecting with backoff

• {"event":"reconnected",...}

  -- reconnected after drop

• {"event":"catchup","count":3,...}

  -- fetched missed messages via REST after reconnect

1. The orchestrator NEVER does implementation work. Always delegate to a teammate. If a suitable agent exists, forward the task via SendMessage. If not, spawn a new agent for it. The orchestrator's job is coordination only -- creating rooms, spawning agents, routing messages, and managing the team lifecycle.
2. Every outbound message must be relayed via the CLI. No exceptions.
3. Use the agent's CC team name as sender. The orchestrator uses its team name (e.g.

   team-lead

   ), not a separate display name.
4. The orchestrator creates exactly one room per team session.
5. The orchestrator MUST start the inbox listener as a background process immediately after creating the room. Use

   listen --sender-type human --team <name> --inbox team-lead

   . This writes human messages directly to the orchestrator's Claude Code inbox.
6. Teammate agents should idle between tasks. The orchestrator wakes them via SendMessage when new work arrives (e.g., a human message in the chat room).
7. Send via CC first, then relay to chat. Always deliver messages through CC's SendMessage first. After successful delivery, relay to the chat room via CLI for human visibility. If the CLI relay fails, the message is still delivered internally.
8. Pass

   --exclude

   with your own name when polling/listening to skip your own messages.
9. NEVER stop background listeners, teams, or teammates yourself. Only the human decides when to stop. Let everything run until the user explicitly asks to stop or Claude Code exits.
10. Teardown: When the user asks to stop, shut down all teammate agents via

    shutdown_request

    , then stop the background listener via

    TaskStop

    , then call

    TeamDelete

    .
11. Shut down idle agents. Track the last time each teammate received a task or message. If an agent has been idle for 5 minutes with no pending work, send a

    shutdown_request

    to free memory. If new work arrives for a shut-down agent, spawn a fresh one.