zhayujie/CowAgent

An open-source personal AI assistant and agent harness that plans tasks, runs tools and skills, and self-evolves with memory and knowledge across many chat channels.

What it is

CowAgent (formerly chatgpt-on-wechat) is a self-hosted Python agent that proactively plans and executes multi-step tasks, controls a computer and external services through tools, and builds a personal knowledge base and long-term memory. It plugs into any major LLM provider and serves multiple messaging channels in parallel from a single instance. It is aimed at individuals and teams who want a 24/7 assistant they can run on a personal computer or server, managed from a unified web console.

Key features

Task planning that decomposes work and loops over tools until the goal is reached.
Three-tier long-term memory (context, daily, MEMORY.md) with a nightly "Deep Dream" distillation pass and hybrid keyword-plus-vector retrieval.
A personal knowledge base that auto-curates structured Markdown with a knowledge-graph view, plus Self-Evolution that reviews conversations to improve skills and consolidate memory.
Built-in tools for file I/O, terminal (bash), browser automation, scheduling, web search, web fetch, vision, and memory, with native MCP integration via mcp.json.
A Skills system with one-click install from Skill Hub, GitHub, or ClawHub, plus conversational skill authoring.
Channels for Web console, Telegram, Slack, Discord, WeChat, Feishu, DingTalk, WeCom, QQ, and more, with multimodal text, image, voice, and file support.

Tech stack

Python 3.7+ per pyproject.toml; the cow CLI is built on Click 8.0+ and requests.
Web console served on port 9899; one-line installer for Linux/macOS, Windows PowerShell, and Docker.
Channel SDKs: python-telegram-bot, slack_bolt, discord.py, wechatpy, lark-oapi (>=1.5.5), dingtalk_stream, websocket-client (>=1.4.0).
Model SDKs: dashscope (Qwen) and zai-sdk (GLM); OpenAI, Claude, Gemini, DeepSeek, and others configured from the console.
Supporting libraries: numpy (>=1.21), aiohttp (>=3.8.6,<3.10), Pillow, croniter, PyYAML, pycryptodome, json-repair, qrcode.
Python 3.13+ support via a GitHub build of web.py plus legacy-cgi.

When to reach for it

You want a self-hosted, always-on personal assistant reachable from Chinese and international messaging platforms alike.
You need multi-step task planning combined with persistent memory and an auto-curated knowledge base.
You want to swap LLM providers freely and route chat, vision, image generation, ASR/TTS, and embeddings to different vendors.
You want to install or author skills and integrate MCP tools without heavy configuration.

When not to reach for it

You want a stateless chatbot; the agent's planning, memory, and evolution add substantial token cost, which the README flags explicitly.
You cannot provide a trusted environment — the agent has access to your local operating system.
You need only a single web chat interface and none of the channel or automation machinery.
You require a non-Python runtime or a fully managed hosted service (offered separately via LinkAI, not this repo).

Maturity signal

Created in August 2022 and pushed within a day of this writing, CowAgent is the most established project of its kind here, with a very large star and fork count and a continuous release cadence (currently the 2.1.x line). Its history as chatgpt-on-wechat, a detailed changelog, and MIT licensing point to a mature, actively maintained codebase that has evolved from a WeChat chatbot into a general agent harness. The large fork count signals heavy real-world reuse.

Alternatives

HKUDS/nanobot — use it when you want a smaller, more minimal self-hosted agent core with a similar channel and MCP surface.
zhayujie/bot-on-anything — use it when you want a lightweight LLM-to-messaging bridge without planning, memory, or a knowledge base (from the same author).
AgentMesh — use it when your problem needs a multi-agent team framework rather than a single personal assistant.

Notes

The project was renamed from chatgpt-on-wechat; the old GitHub URL redirects, and existing users can optionally update their git remote. Despite the WeChat origin, requires-python in pyproject.toml is still 3.7+ while the requirements pin newer libraries and add Python 3.13 workarounds, so the effective supported range is narrower than the manifest alone suggests.

Wiki: zhayujie/CowAgent