MMClaw

MMClaw is an ultra-lightweight, pure Python multimodal AI Agent kernel. Following the "Kernel, Not App" philosophy, it requires zero compilation and zero runtime dependencies (except for the WhatsApp bridge), making it suitable for environments ranging from Raspberry Pi to servers. Current version 0.0.85, GPL-3.0 licensed, Python >= 3.8 required.

Core Capabilities#

• Multi-Channel Adapters: Built-in connectors for Terminal, Telegram, WhatsApp, WeChat, Feishu, and QQ Bot. No public IP or webhook needed.
• Broad LLM Support: Adapts to OpenAI (incl. Codex), Gemini (incl. Vertex AI, Gemini CLI), DeepSeek, Kimi, MiniMax, Claude, OpenRouter, and any OpenAI-compatible endpoints (Ollama, LocalAI, etc.).
• Multimodal & File Processing: Supports uploading images, PDFs, and documents via chat; Agent can read, analyze, and act accordingly.
• Intelligent Skill System: Introduces SkillKG (Skill Knowledge Graph) for dependency reasoning and safety checks, with extensible skills manageable via mmclaw skill install/uninstall.
• Automation Orchestration: Combines optional Playwright browser automation, built-in Web search, and APScheduler-based cron scheduling for complex workflows.
• State Management: Features cross-session persistent memory and fully isolated multi-workspace execution. All data stored in ~/.mmclaw by default.
• A2A Communication: Natively integrates ClawMeets platform. Each account identified by a 12-character public address with file attachment, contact management, and auto-notification support.
• Non-Interactive Execution: mmclaw run -p "prompt" runs a complete agentic loop and exits, easy to integrate into external script pipelines.
• AI CLI Coding Collaboration: Can drive Codex, Gemini CLI, Claude Code and other coding sessions.

Architecture Design#

Adopts a highly centralized single-layer modular design with core modules:

• kernel.py: Agent kernel with Agent loop and tool-calling logic
• providers.py: LLM Provider adaptation layer
• connectors.py: Multi-channel connector implementations
• tools.py: Built-in tools (Shell execution, Web search, file ops, browser automation, etc.)
• memory.py: Persistent memory management
• config.py: Configuration management center
• watcher.py: Monitoring and scheduled task dispatch
• skills/ & skill-kg/: Skill packages and Skill Knowledge Graph data

Except for the WhatsApp connector depending on bridge.js (Node.js), the rest is 100% pure Python.

Installation & Usage#

pip install mmclaw
mmclaw run
# For Feishu connector
pip install "mmclaw[all]"

After first run, configure connectors and LLM providers via mmclaw config.

Unconfirmed Items#

• ClawMeets platform independence and server-side implementation details
• Playwright integration current status (commented out in setup.py)
• Claude Provider implementation method (possibly via OpenAI-compatible endpoint)
• ClawHub skill distribution platform existence (no public URL)
• No published GitHub Releases found; version management approach unconfirmed