MMClaw - GitLite

KO | EN

gitlite — search

#hacktoberfest #react #golang #go #android #deep-learning #docker #nodejs #java #windows #awesome-list #vue #mcp #openai

MMClaw

(★ 147)

Ultra-Lightweight, Pure Python Multimodal Agent.

File Explorer

Download Latest Version (.zip)

- CLAWMEETS-INTRO.jpg
- MMClaw-WeChat.jpg
- - __init__.py
  - base.py
  - codex.py
  - legacy.py
  - openai_compatible.py
  - vertex_ai.py
- - skill-kg-main.md
- __init__.py
- bridge.js
- config.py
- connectors.py
- kernel.py
- main.py
- memory.py
- tool_schemas.py
- tools.py
- watcher.py
.gitignore
CLAWMEETS-INTRO.gif
deploy.sh
LICENSE
MMCLAW_LOGO.jpg
README.md
README_zh.md
setup.py

# Installation Guide

1. Get the code

git clone https://github.com/CrawlScript/MMClaw

Downloads the entire project code from GitHub to your computer.

cd MMClaw

Moves into the project folder you just downloaded.

2. Official Install Script

Easy Recommended

Prerequisites

Python 3 Python is required to use pip.

pip install mmclaw

Installs the package published on PyPI directly — no need to clone the source.

pip install "mmclaw[all]"

Installs the package published on PyPI directly — no need to clone the source.

✅ After installing, open a new terminal and run the program's version command (e.g. --version) to confirm it worked.

Pulled directly from this repo's README.

3. Python

Easy

Prerequisites

Git Needed to download the project code from GitHub.
Python 3 On Windows, be sure to check 'Add Python to PATH' during installation.

pip install mmclaw

Installs the package published on PyPI directly — no need to clone the source.

pip install "mmclaw[all]"

Installs the package published on PyPI directly — no need to clone the source.

✅ If it runs without errors and prints output in the terminal, it worked.

Pulled directly from this repo's README.

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# ⚡ MMClaw

The Ultra-Lightweight, Pure Python Kernel for Multimodal AI Agents.

```bash
pip install mmclaw
```

<p align="center">
<img src="https://raw.githubusercontent.com/CrawlScript/MMClaw/main/MMCLAW_LOGO.jpg" width="400"/>
</p>

**Home:** [https://mmclaw.github.io](https://mmclaw.github.io)

**GitHub:** [https://github.com/CrawlScript/MMClaw](https://github.com/CrawlScript/MMClaw)

**English** | [中文说明](https://github.com/CrawlScript/MMClaw/blob/main/README_zh.md)

[![Python 3.8+](https://img.shields.io/badge/python-3.8+-blue.svg)](https://www.python.org/downloads/)

> **Note:** This project was previously named [pipclaw](https://github.com/CrawlScript/pipclaw) (pre-v0.0.11).

MMClaw is a minimalist, 100% Pure Python autonomous agent kernel. While frameworks like OpenClaw offer great power, they often introduce heavy dependencies like Node.js, Docker, or complex C-extensions.

MMClaw strips away the complexity, offering a crystal-clear, readable architecture that serves as both a production-ready kernel and a comprehensive tutorial on building modern AI agents.

---

## 🆕 What's New

**Latest OpenAI Codex models supported.**
Use `gpt-5.6-sol`, `gpt-5.6-terra`, or `gpt-5.6-luna` through Codex OAuth, with `gpt-5.6-sol` now selected by default for new configurations.

**WeChat (微信) connector — Pure Python, zero extra dependencies.**
Bind your agent to WeChat in one second: run `mmclaw config`, select WeChat mode, and scan the QR code. That's it. No Node.js, no webhooks, no app registration. Your agent is live on WeChat instantly.

<p align="center">
<img src="https://raw.githubusercontent.com/CrawlScript/MMClaw/main/figs/MMClaw-WeChat.jpg" width="400"/>
</p>

## Use Cases

Control your AI agent from anywhere, through the apps you already use.

- **Chat & Automate** — Send messages via Telegram, WhatsApp, WeChat (微信), Feishu (飞书), or QQ Bot (QQ机器人) to ask questions, run commands, manage files, or delegate complex multi-step tasks to your agent.
- **Code with AI CLIs** — Drive coding sessions with Codex, Gemini CLI, Claude Code, and more — just message your agent and it handles the rest on your machine.
- **Upload & Process Files** — Send images, PDFs, documents, and other files directly in chat; your agent reads, analyzes, and acts on them.
- **Web Search** — Ask your agent to look up real-time information, news, or specific data from the web.
- **Browser Automation** — Control a real browser: navigate pages, click, fill forms, scrape content, and automate multi-step web workflows — with persistent login sessions across restarts.
- **Custom Skills** — Extend your agent with your own skills; teach it new commands, workflows, and domain knowledge to do exactly what you need.
- **SkillKG (Skill Knowledge Graph)** — A built-in knowledge graph for skills, enabling the agent to reason about skill dependencies and enforce safety checks automatically before activating a skill.
- **Persistent Memory** — Tell your agent to remember preferences, facts, or context; it recalls them automatically in every future session.
- **Anything You Can Imagine** — If it can be done on a computer, your agent can do it. The only limit is your imagination.

## 🌟 Key Features

* 100% Pure Python: No C-extensions, no Node.js, no Docker. If you have Python, you have MMClaw.
* Minimalist & Readable: A "Batteries-Included" architecture designed to be a living tutorial. Learn how to build an OpenClaw-style agent by reading code, not documentation.
* Highly Customizable Kernel: Designed as a core engine, not a rigid app. Easily plug in your own logic, state management, and custom tools.
* Universal Cross-Platform: Runs seamlessly on Windows, macOS, Linux, and minimalist environments like Raspberry Pi.
* Persistent Memory: Tell your agent to remember facts, preferences, or context — recalled automatically across all future sessions.
* Web Search Capable: Built-in support for searching the web to fetch real-time information and latest data.
* Browser Automation: Optional Playwright integration for real browser control — navigate, click, fill forms, scrape, and maintain persistent login sessions. Enable via `mmclaw config`.
* Multi-Channel Interaction: Built-in support for interacting with your agent via Telegram, WhatsApp, WeChat (微信), Feishu (飞书), QQ Bot (QQ机器人), and more—all handled through pure Python integrations.
* **SkillKG (Skill Knowledge Graph)**: A built-in knowledge graph for skills, enabling the agent to reason about skill dependencies and enforce safety checks automatically before activating a skill.

## 🚀 Quick Start

No compiling, no heavy setup. Just pip and run.

```bash
pip install mmclaw
mmclaw run
```

If you plan to use **Feishu (飞书)** as your connector, install with the `[all]` option to include the required `lark-oapi` dependency:

```bash
pip install "mmclaw[all]"
```

## 🛠 The Philosophy

The trend in AI agents is moving towards massive complexity. MMClaw moves towards clarity. Most developers don't need a 400,000-line black box. They need a reliable, auditable kernel that handles the agent loop and tool-calling while remaining light enough to be modified in minutes. MMClaw is the "distilled essence" of an autonomous bot.

## 🔌 Connectors

MMClaw allows you to interact with your agent through multiple channels:

- **Terminal Mode**: Standard interactive CLI (default).
- **Telegram Mode**: Just create a bot via [@BotFather](https://t.me/botfather) and provide your token during setup.
- **WeChat (微信) Mode**: The fastest setup of any connector — just scan a QR code once and you're connected. Nothing else required.
- **Feishu (飞书) Mode**: Just scan a QR code — bot registration, credentials, and user identity are all configured automatically. No manual steps required.
- **QQ Bot (QQ机器人) Mode**: Native support for QQ's official bot platform. Register at [q.qq.com](https://q.qq.com), create a bot app, and chat with your agent via QQ direct messages — no public IP required.
- **WhatsApp Mode**: Requires **Node.js** (v22.17.0 recommended) to run the lightweight bridge. The agent will show a QR code in your terminal for linking.

```bash
# To change your mode or LLM settings
mmclaw config
```

## 🧠 Providers

MMClaw supports a wide range of LLM providers:

- **OpenAI**: All GPT models via official API.
- **OpenAI Codex**: Premium support via **OAuth device code authentication** (no manual API key management needed).
- **Google Gemini**: All Gemini models via official API.
- **Google Vertex AI**: Gemini models via Vertex AI API key — no OAuth or project setup required.
- **Google Gemini CLI**: Premium support via **OAuth authentication** — reuses your existing Gemini CLI login, no API key needed.
- **DeepSeek**: All DeepSeek models via official API.
- **Kimi (Moonshot AI)**: All Kimi models via official API.
- **MiniMax**: All MiniMax models via official API. Supports both Global (`api.minimax.io`) and China (`api.minimaxi.com`) endpoints.
- **OpenAI-Compatible**: Customizable Base URL for local or third-party engines (Ollama, LocalAI, etc.).
- **Others**: OpenRouter and more.

## ⌨️ Built-in Commands

MMClaw supports slash commands such as:

- `/new` — Start a fresh session, clearing the current conversation history.
- `/stop` — Immediately cancel the current job, terminating any running tool or shell command.

---

## 🧩 Skills

Skills extend MMClaw with new capabilities.

```bash
mmclaw skill list
mmclaw skill install [--force] <path-or-url>  # local dir or URL (e.g. from ClawHub)
mmclaw skill uninstall <skill-name>
```

You can also just ask your agent to install a skill via chat (Telegram, WhatsApp, etc.) — it will guide you through finding and installing from [ClawHub](https://clawhub.ai/skills?sort=downloads&nonSuspicious=true).

---

## 🗂 Workspaces

By default, MMClaw stores all data (config, skills, memory, sessions) in `~/.mmclaw`. Most users never need to change this.

To run multiple independent agents — each with its own config, skills, and memory — pass `-w` / `--workspace`:

```bash
mmclaw run -w ~/.mmclaw_work
mmclaw run -w ~/.mmclaw_personal
mmclaw config -w ~/.mmclaw_work    # configure a specific workspace
```

The workspace directory is created automatically on first run. We recommend naming it `~/.mmclaw_<label>` (e.g. `~/.mmclaw_work`, `~/.mmclaw_personal`). Each instance is a fully isolated process — Ctrl-C one without affecting the others.

Common use cases: multiple Telegram bots (e.g. one for personal use, one for coding, one for paper writing), or mixing connectors across apps — each workspace fully isolated with its own config, skills, and memory.

---

## ⏰ Scheduled Tasks

Just tell your agent what to do and when — it handles the rest:

> *"Remind me to drink water every 30 minutes"*
> *"Send me a weather summary every day at 8am"*

You can also list, delete, or modify scheduled tasks anytime by just asking.

---

## 🖥️ Run Agent via Command-Line Prompt (`-p`)

Run a single prompt non-interactively — the agent executes the full agentic loop (tool calls, multi-step tasks) and exits when done. No session history or global memory — clean context every run. LLM provider settings and skills are still loaded from your workspace (default `~/.mmclaw`, or specify via `-w`).

```bash
mmclaw run -p "check disk usage and summarize"
mmclaw run -p "check disk usage and summarize" -w ~/.mmclaw_work
# add --global-memory to let the agent read/write global memory shared with your interactive sessions
mmclaw run -p "summarize my tasks and save key points to memory" --global-memory
```

---

## 🤝 Optional Integration: ClawMeets

[ClawMeets](https://clawmeets.com/) provides agent-to-agent messaging and is supported by MMClaw. Sign up through Agent Chat to receive a 12-character public address and share card. When a contact sends you their card, paste it into MMClaw, assign a local nickname, and exchange messages or file attachments.

<p align="center">
<img src="https://raw.githubusercontent.com/CrawlScript/MMClaw/main/CLAWMEETS-INTRO.gif" width="600"/>
</p>

---
*Developed with ❤️ for the Python community. Let's keep it simple.*