CowAgent

CowAgent is an open-source super AI assistant and reference implementation of Agent Harness engineering, published under the MIT License by zhayujie. It plans tasks, runs built-in tools and installable skills, keeps long-term memory and a personal knowledge base, and can be reached through a web console and many chat channels. The latest release listed is v2.2.1, published 2026-10-14.

What It Is

CowAgent is a self-hosted AI agent that runs on a personal computer or server around the clock. Messages arrive through channels, the Agent Core plans and reasons over memory, knowledge, tools and skills, and a configured model generates the response, which is returned through the originating channel. Each layer is decoupled and independently extensible.

It can be installed with a one-line script on Linux, macOS or Windows, or via Docker. Once running, a web console at port 9899 is used to chat with the agent, configure models, connect channels and install skills. A desktop client for macOS and Windows bundles the backend.

Tools, Skills and MCP

Built-in tools cover file I/O, terminal, file sending, memory retrieval, environment variables, web fetching, scheduling, web search, vision and browser automation. Skills are higher-level workflows defined by a manifest that compose tools; they can be installed from Skill Hub, GitHub, ClawHub or a URL, or created through natural-language conversation with a skill-creator. MCP servers are integrated via a single mcp.json, supporting stdio and SSE transports and hot reload.

Memory, Knowledge and Self-Evolution

Memory uses three tiers (conversation context, daily memory, and a long-term MEMORY.md), with a nightly Deep Dream pass that distills memories into long-term entries, and hybrid keyword plus vector retrieval. The knowledge base organizes information by topic into a Markdown wiki with cross-references and a visual knowledge-graph view. Self-Evolution reviews conversations to improve skills, follow up on unfinished tasks and consolidate memory.

Models, Channels and Teams

Chat, vision, image generation, speech recognition, text-to-speech and embeddings can each be routed to different providers, including OpenAI, Claude, Gemini, DeepSeek, Qwen, GLM, Kimi, MiniMax, Doubao, ERNIE, MiMo, LinkAI and custom endpoints. Channels include Web, Telegram, Slack, Discord, WeChat, Feishu/Lark, DingTalk, WeCom and QQ. Multi-agent teams let several agents with their own roles, models, skills and knowledge collaborate in a shared conversation, with sub-agents for parallel delegation.

Tradeoffs to Know

The project notes that agent mode consumes substantially more tokens than regular chat and that the agent has access to the local operating system, so it should only be deployed in trusted environments. It was previously named chatgpt-on-wechat.