CowAgent — an open-source AI agent harness
zhayujie/CowAgentCowAgent is an open-source AI assistant and agent harness (formerly chatgpt-on-wechat) that plans tasks, runs tools and skills, and maintains memory and a knowledge base across many chat channels.
A self-hostable, multi-channel AI assistant for people who want one agent running across web and messaging apps with pluggable LLMs. Good for tinkerers and small teams; the breadth of features means more to configure and vet than a single-purpose bot.
The problem it solves
Most chatbots are tied to one model and one channel, forget context between sessions, and can't take real actions. Wiring up planning, memory, tools and several messaging platforms yourself is a large amount of glue code.
What is it?
CowAgent is an open-source assistant written in Python that the project describes as a reference implementation of an 'agent harness'. Messages arrive through channels, an agent core plans over memory, knowledge, tools and skills, and a pluggable model generates the reply that is sent back through the same channel. It ships as a one-line installer with a unified web console and supports multiple LLM providers and messaging platforms.
Why it's getting attention
CowAgent (formerly chatgpt-on-wechat) sits in the active AI-agent space and has roughly 46,000 GitHub stars. Its draw is packaging planning, memory, skills and many messaging channels into one self-hostable assistant that installs in a single command and swaps model providers from a web console.
Key features
- ✓Task planning that decomposes goals and loops over tools until done
- ✓Tiered memory (context, daily, core) with keyword + vector retrieval
- ✓Knowledge base curated into a Markdown wiki and knowledge graph
- ✓Skills installable from a Skill Hub, GitHub or created via conversation
- ✓Built-in tools: file I/O, terminal, browser, scheduler, web search and native MCP
- ✓Multi-channel: Web, WeChat, Feishu, DingTalk, WeCom, QQ, Telegram and Slack
- ✓Multi-model support (Claude, GPT, Gemini, DeepSeek, Qwen and others), switchable from the console
Best use cases
- •Running one assistant across web and messaging platforms for a person or small team
- •Automating multi-step tasks with tools, scheduling and MCP integrations
- •Building a personal knowledge base and long-term memory from ongoing chats
- •Prototyping an agent without committing to a single LLM vendor
How to install / try
The README documents a one-line installer that handles dependencies, config and startup. Linux/macOS: `bash <(curl -fsSL https://cdn.link-ai.tech/code/cow/run.sh)`. Windows (PowerShell): `irm https://cdn.link-ai.tech/code/cow/run.ps1 | iex`. Docker: download the compose file and run `docker compose up -d`.
How to use
Once started, open the web console at `http://localhost:9899` to chat with the agent, configure models, connect channels and install skills. For server deployment the README notes setting `web_host` to `0.0.0.0` and a `web_password` in `config.json`, and opening port 9899. More is documented at docs.cowagent.ai.
Strengths
- ✓One-line install and a single web console lower setup friction
- ✓Model-agnostic: swap between many LLM providers from the UI
- ✓Broad channel coverage out of the box
- ✓MIT-licensed and actively documented
Limitations & risks
- △The one-line installer pipes a remote script into bash/PowerShell — review it before running
- △Install and skill hubs depend on third-party hosted endpoints (link-ai.tech)
- △Much of the detail lives in external docs (docs.cowagent.ai), not the README
- △The broad feature set (memory, knowledge, evolution, many channels) means more to configure and secure
Alternatives
Who should try it — and who should skip
Choose CowAgent if you want a self-hosted assistant that spans several messaging platforms and lets you swap LLM providers freely, and you're willing to vet a one-line installer and configure it. Skip it if you need a narrowly-scoped, audited bot or a fully managed SaaS assistant.
Frequently asked questions
The README states CowAgent was formerly named chatgpt-on-wechat; it is the continuation of that project under a new name and broader scope.
The README lists Claude, GPT, Gemini, DeepSeek, Qwen, GLM, Kimi, MiniMax, Doubao and more, switchable from the web console.
Via a documented one-line installer for Linux/macOS and Windows, or with Docker Compose. After startup the web console runs on port 9899.
The repository is MIT licensed.