Skip to main content

Multi-Chat Permissions

MicroClaw uses a chat-scoped authorization model for tool calls.

Model​

  • Non-control chats can only operate on their own chat_id
  • Control chats can operate across chats
  • Global memory writes (write_memory with scope: "global") require control-chat privileges

Enforced tools:

  • send_message
  • Scheduler tools (schedule_task, list_scheduled_tasks, pause_scheduled_task, resume_scheduled_task, cancel_scheduled_task, get_task_history)
  • export_chat
  • todo_read / todo_write
  • chat-scoped memory operations (read_memory/write_memory with scope: "chat")

Configure control chats​

Add control_chat_ids to microclaw.config.yaml:

control_chat_ids: [123456789, 987654321]

Then restart:

microclaw start

Find chat IDs​

If chats already have messages stored, query SQLite:

sqlite3 ~/.microclaw/runtime/microclaw.db \
  "SELECT chat_id, chat_title, chat_type, last_message_time FROM chats ORDER BY last_message_time DESC;"

Use the chat_id values from this output in control_chat_ids.

Quick verification​

From a non-control chat:

  1. Ask the bot to send a message to a different chat_id
  2. Expected: permission error

From a control chat:

  1. Repeat the same cross-chat request
  2. Expected: permission layer allows execution (final success depends on target chat reachability/tool behavior)

Notes​

  • This is authorization at tool execution time, using trusted caller context from runtime.
  • It does not sandbox host command/file access by itself. Keep host deployment locked down.
  • To inspect sandbox posture and execution policy state in Web UI, open Settings and run config self-check (Security posture card).