CowAgent

One line: CowAgent (formerly chatgpt-on-wechat, 47k+ stars) is a popular open-source personal AI assistant from the Chinese dev community: install the desktop client and you are running — a single web console for chatting, connecting channels (WeChat, Feishu, Telegram…), and installing skills, with whatever model you like.

What you get

• A ready-to-run desktop client (runtime bundled — double-click, no Python required)
• One web console (http://localhost:9899) for chat, model setup, channels and skills
• Multi-channel service in parallel + Skill Hub marketplace + scheduled tasks / memory / knowledge base

Before you start (2 minutes)

• Windows 10/11 x64 or macOS (Intel and Apple Silicon installers both published)
• One model key (DeepSeek, Claude, OpenAI, Gemini, Qwen, GLM, Kimi…) — or the CowAgent aggregated key service
• Not needed: administrator rights (installs into your user profile), Docker, Python

Advanced users who want the headless service mode (parked on a server, serving channels like WeChat) can use the manual route below — that one needs Python 3.7+.

Install

Route A: install with AiBOX (recommended)

1. Open AiBOX → app directory → CowAgent → Install
2. AiBOX downloads the official installers (sha256-verified, accelerated automatically in CN): silent NSIS install to %LOCALAPPDATA%\Programs\CowAgent on Windows, dmg mount + copy on macOS — no admin rights at any point
3. Launch CowAgent from the Start menu / Launchpad

Route B: manual install (official channels)

• Desktop client: download the installer for your OS from cowagent.ai/download and run it
• One-line service install (Linux / macOS, lands in a CowAgent folder in your current directory):

bash <(curl -fsSL https://cdn.link-ai.tech/code/cow/run.sh)

Windows PowerShell service install:

irm https://cdn.link-ai.tech/code/cow/run.ps1 | iex

Then open http://localhost:9899 in your browser (service mode).

Configuration (connect a model + channels)

1. Open the console → Model settings: pick a provider from the table (DeepSeek / Claude / OpenAI / Gemini / Qwen…) and save your key — everything happens in the UI, no config file editing
2. Prefer a single key for everything? Choose the CowAgent aggregated model service
3. Free local route: install Ollama, then point the Custom entry at your local endpoint
4. Channels: follow the in-console wizard for Telegram / Discord / WeChat customer service; skills: install one from Skill Hub in a click

Verification: what "success" looks like

1. The client opens normally and the console at http://localhost:9899 is reachable
2. Send "hello" in the chat box → a normal reply means the model is connected
3. Service route: cow status shows running; cow logs has no errors
4. Stronger check: install a Skill Hub skill and see it listed by /skill list in the conversation

FAQ

• Windows says the publisher is unknown: the official installer is signed; if blocked, use "More info → Run anyway", or install via the AiBOX route
• Port 9899 already in use: stop the old instance with cow stop, or change the port in the config
• Domestic channels (WeChat etc.) won't connect: usually network or an expired QR code — re-run the channel wizard in the console
• Upgrading: desktop → download and overwrite-install the new version; service → cow update
• Formerly chatgpt-on-wechat: the repo is now CowAgent; old links redirect automatically (git remote set-url updates your clone)

Uninstall

• Desktop client: Windows Settings → Apps → CowAgent → Uninstall; macOS: move CowAgent.app to the Trash
• Service: cow stop, then delete the CowAgent folder
• ⚠️ Chat history, model keys and installed skills live in the app data folder (%APPDATA%\CowAgent / ~/Library/Application Support/CowAgent); the uninstaller does not delete them — back up first if you clean them manually

More: quick start · CLI reference · Skill Hub