// the find
zhayujie/CowAgent
Open-source super AI assistant & Agent Harness. Plans tasks, runs tools and skills, self-evolves with memory and knowledge. Multi-model, multi-channel. Lightweight, extensible, one-line install. (formerly chatgpt-on-wechat)
CowAgent is a Python-based personal AI agent that connects to LLM providers and delivers responses through a wide range of messaging platforms — WeChat, Telegram, Slack, DingTalk, and a built-in web console. It evolved from the popular chatgpt-on-wechat project and now adds multi-step task planning, a three-tier memory system, a skills marketplace, and MCP tool integration. The target user is someone who wants a self-hosted, always-on AI assistant they can reach through whatever chat app they already live in.
The channel coverage is genuinely impressive — twelve platforms with consistent feature matrices for text, image, file, and voice, handled through a clean factory pattern rather than spaghetti conditionals. The three-tier memory architecture (context → daily → core with a nightly distillation pass) is a thoughtful take on the context window problem that most hobby agent projects just ignore. MCP integration via a single mcp.json with hot reload means you get access to the whole MCP ecosystem without writing glue code. The release cadence over the last eight weeks (a tagged release roughly every two weeks with real feature additions) suggests this is actively maintained, not abandoned.
The one-line installer pipes a remote shell script directly to bash with no integrity check — that is a supply chain attack waiting to happen, and for a tool that has shell access to your machine, it matters. The skills system relies on a proprietary Skill Hub backed by LinkAI, which is a commercial entity; if that service goes away or changes terms, the ecosystem collapses. The memory architecture stores everything as Markdown files rather than a queryable store, which will degrade as context grows large — the 'Deep Dream' distillation is a band-aid, not a solution. There is no multi-user model: the web console is password-protected but single-tenant, so sharing the instance between household members or a small team requires workarounds that are not documented.