A super AI agent framework powered by LLMs, supporting multi-channel access (WeChat/Feishu/DingTalk/etc.) with autonomous task planning, long-term memory, knowledge base, and skills system
CowAgent (formerly chatgpt-on-wechat) is a super AI agent powered by large language models, combining out-of-the-box AI assistant capabilities with a highly extensible agent framework. It supports multi-channel access through WeChat, Feishu, DingTalk, QQ, WeCom, Web, and Terminal, with multiple channels running simultaneously.
On the agent capability side, CowAgent features autonomous task planning and a decision engine that understands complex tasks and continuously invokes tools and skills until completion, with intelligent context compression and a deep thinking mode (enable_thinking config, off by default). The long-term memory system automatically persists conversation memories to local files and databases, covering core memory, daily memory, and dream distillation mechanisms, with keyword and vector retrieval support. The personal knowledge base automatically organizes structured knowledge and builds knowledge graphs through cross-referencing.
The Skills system supports one-click installation from Skill Hub or GitHub, as well as creating skills through natural language conversation. Built-in tools cover file read/write, terminal execution, browser operations (Playwright + Chromium), scheduled tasks, web search, and more — all autonomously invoked by the agent.
Multi-model support covers OpenAI, Claude, Gemini, DeepSeek, MiniMax, GLM, Qwen, Kimi, Doubao, Azure OpenAI, Baidu Wenxin, iFlytek Xinghuo, and LinkAI, with MiniMax-M2.7 as the default model. Multimodal capabilities support parsing and generation of text, images, voice, and files.
The project provides a built-in Web console (default port 9899) and Cow CLI for managing models, skills, memory, channels, scheduled tasks, and logs. It supports one-click script installation, source code installation, and Docker deployment, with a deployment-free online experience available via the LinkAI platform. Language composition: Python 84.5%, JavaScript 8.3%, HTML 3.3%, Shell 1.7%, CSS 1.3%, PowerShell 0.9%. Supports Python 3.7–3.13 on Linux/macOS/Windows.
Architecture Layers:
agent/: Agent core logic (task planning, decision engine, memory management)bridge/: Bridge layer (message routing and transformation between channels and models)channel/: Channel access layer (message send/receive implementations per platform)models/: Model interface adaptation layer (LLM API wrappers per vendor)skills/: Skills systemplugins/: Plugin systemvoice/: Voice processing moduletranslate/: Translation modulecli/: Command-line tool (Cow CLI)
Message Flow: channel/ (receive) → bridge/ (route & transform) → agent/ + models/ (reasoning & decision) → bridge/ → channel/ (reply)
Typical Scenarios: Personal AI assistant (24/7 via IM tools), enterprise digital employee (smart customer service, private domain operations, efficiency assistant — covering consumer, health, education, and tech manufacturing industries), agent development framework (extending model interfaces, channels, tools, and skills).
Unconfirmed Items: The specific technical approach for WeChat integration (web protocol / iPad protocol / hook) requires inspecting channel/weixin/ source code; dream distillation memory algorithm details are not expanded in the README; Skills system format specifications and Skill Hub listing process require consulting the documentation center; Docker image registry address and versioning strategy are not explicitly listed (image name retains chatgpt-on-wechat); the boundary between the plugin system (plugins/) and skills system (skills/) is not clearly distinguished.