Description
CowAgent is an open-source project, previously named chatgpt-on-wechat and now officially CowAgent, describing itself as a "super AI assistant" and a reference implementation of Agent Harness engineering. Rather than a single chatbot, it's built to proactively plan tasks, control a computer and outside services, create and run its own Skills, maintain a personal knowledge base and long-term memory, and run multiple agents as a coordinated team, with a self-evolution process that improves it through use.
Read More
The architecture is straightforward: messages come in through a Channel, such as the web console, WeChat, Telegram, or Slack, the Agent Core plans and reasons using memory, knowledge, and available tools and skills, a connected model provider generates the response, and the reply goes back out through the same channel. Memory is organized in three tiers, context, daily, and core, with automatic distillation and a hybrid of keyword and vector retrieval for recall. Knowledge is curated automatically into a structured Markdown wiki and an evolving, visually browsable knowledge graph. The self-evolution process reviews past conversations to improve skills, follow up on unfinished tasks, and consolidate memory and knowledge over time.
Skills can be installed with one click from CowAgent's own Skill Hub, from GitHub, or from a service called ClawHub, or built from scratch by describing what's needed in natural language. Built-in tools cover file I/O, a terminal, browser automation, a scheduler, memory retrieval, and web search, among others, with native support for the Model Context Protocol so it can connect to external tools. Multiple agents can be organized into a team, each with its own role, model, skills, and knowledge, collaborating in a shared conversation. CowAgent works with text, images, voice, and files, and can be connected to several mainstream LLM providers, switchable from the web console without editing configuration files.
Getting started is a one-line installer for Linux, macOS, or Windows, or a Docker Compose setup; either way, a local web console opens for chatting, configuring models, connecting channels, and installing skills. A bundled desktop client is also available for macOS and Windows. Day-to-day operation goes through a command-line tool for starting, stopping, updating, and checking logs. The project is released under the MIT license and is aimed at developers and power users who want a self-hosted, extensible AI assistant rather than a closed, single-vendor product.
- CategoryAI Agent Builders
- Pricing ModelOpen source
- DeploymentSelf-hosted
- PlatformsLinux, macOS, Windows
CowAgent's Core Features
Plans and executes multi-step tasks by looping over available tools until the goal is reached
Multi-agent teams where each agent has its own role, model, skills, and knowledge
Three-tier memory (context, daily, core) with hybrid keyword and vector retrieval
Auto-curates conversations into a structured knowledge base and an evolving knowledge graph
Self-evolution process that reviews past conversations to improve skills and follow up on unfinished tasks
One-click skill installation from a Skill Hub, GitHub, or ClawHub, plus custom skills built via natural language
Built-in tools for file I/O, terminal, browser automation, scheduling, and web search, with native MCP support
Works across a web console and messaging channels including WeChat, Telegram, Slack, Feishu, and DingTalk
Switch between multiple LLM providers from the web console without editing config files
Getting Started with CowAgent
Install: Run the one-line installer for Linux, macOS, or Windows, or start the Docker Compose file
Open the console: Open the local web console the installer starts
Connect a model: Add an LLM provider's API key from the console
Connect a channel: Link WeChat, Telegram, Slack, or another supported channel if you want it beyond the web console
Install or build skills: Add skills from the Skill Hub or create one by describing it in natural language
Manage the service: Use the command-line tool to start, stop, update, or check logs
CowAgent's Use Cases
- Self-hosted AI assistant
- Multi-agent workflows
- Messaging bot deployment
