Skip to main content

FAQ

Is MicroClaw production-ready?​

MicroClaw is under active development. It is stable enough for personal and small-team use, but you should review the security model and run it on a locked-down host.

Does it support images, voice, or files?​

Telegram supports text + images + document uploads. Voice messages are transcribed when openai_api_key is configured.
For tool-driven outbound attachments, send_message supports Telegram / Discord via attachment_path (+ optional caption).

Can I restrict who can run commands?​

Yes. Configure control_chat_ids in microclaw.config.yaml:

  • non-control chats can only operate on their own chat_id
  • control chats can perform cross-chat actions
  • global memory writes are limited to control chats

How does memory work?​

Memory is stored in AGENTS.md files under ~/.microclaw/runtime/groups/ by default (or under your configured data_dir) and injected into the system prompt for every request. There are three scopes: global (groups/AGENTS.md), bot/account (groups/{channel}/AGENTS.md), and chat (groups/{channel}/{chat_id}/AGENTS.md).

How do scheduled tasks work?​

The scheduler polls every 60 seconds for due tasks, runs the same agent loop as normal messages, sends the result, and updates the next run time.

Can I add custom tools?​

Yes. Implement the Tool trait (microclaw_tools::runtime::Tool) in src/tools/, register it in ToolRegistry::new(), and it becomes available to the LLM automatically.

What model is used by default?​

model defaults to a provider-specific value (claude-sonnet-4-5-20250929 for anthropic, gpt-5.2 for openai, gpt-5.3-codex for openai-codex, llama3.2 for ollama).
You can override it in microclaw.config.yaml.