CowAgent

You will like this if your day happens inside WeChat, Feishu or DingTalk and you want an assistant that answers there instead of in another browser tab. The daily trait is that one instance serves every connected channel in parallel, and channels are onboarded from the web console, so the second one takes minutes.

Where it hurts is that it is a chat assistant first and a coding tool a distant second: no git operations, no multi-file edits. Pick it for a personal assistant that lives in Chinese chat apps. Pick QwenPaw if you want the same channels with a sandbox under the shell.

The risk is stated by the maintainers. The disclaimer says the agent has access to your local operating system and should only be deployed in trusted environments. The recommended install pipes a script from cdn.link-ai.tech into bash, and the Docker path fetches its compose file from the same CDN. A harness with a bash tool, browser automation and a scheduler runs with your user's permissions on a box you were told to trust.

The consequence: run the Docker path on a machine that holds nothing you value. What it does right: the console binds to localhost by default, and exposing it requires setting web_host and a web_password on purpose.

The documented loop is decomposition, then tool calls until the goal is reached, which is the standard shape. What is distinctive is memory. 1. Short-term conversation context. 2. Mid-term daily memory. 3. Long-term MEMORY.md, into which a nightly Deep Dream pass distills the day. Retrieval is hybrid keyword and vector search, and a separate knowledge base is curated by topic into a Markdown wiki. Self-Evolution reviews past conversations to revise skills.

No benchmark is published, so how much of the planning claim is demonstrated cannot be assessed, and verification of edits is not described. The observation: a nightly distillation step assumes the assistant is left running overnight.

An individual maintainer's repository with a company at the end of the README. LinkAI, reached through a simple-future.tech mailbox or a WeChat QR code, sells the managed runtime, workspaces, role-based access, audit logs and private deployment. The rename from chatgpt-on-wechat is the tell: a brand built on one messaging platform is being repositioned as a general harness, and the star count came along for the ride.

Moat: the Chinese channel integrations, which Western projects rarely build. Pricing power: none on the open project, some on the hosted one. Likely acquirer: a Chinese cloud that wants the install base, or nobody. Position: long the channels, neutral on the company.

The demo is an assistant that answers in a group chat. For sixty engineers the software costs nothing and the meter is whichever model keys we hand it. Chat, vision, image generation, speech and embeddings can each route to a different vendor, so the data-retention review is five reviews, not one. Nothing in the open project offers SSO or SCIM; those are sold in LinkAI, a separate product with its own contract.

CI fit is nil; this is a chat assistant, not a build step. Onboarding is easy for an engineer and support is a Discord server plus GitHub issues. Not yet.

MIT and Python, with the extension points where I expect them. MCP is a single mcp.json with stdio and SSE transports and hot reload, so tools appear without a restart. Models are a provider table with a Custom row for a local server or a third-party proxy, and the Skill Hub is not the only source: /skill install takes GitHub and ClawHub too, and a skill is a manifest plus tools I can read.

What I resent is the cow CLI wrapping everything in a service I did not ask for. Still, it is readable end to end, and a fork keeps every channel adapter. Enough for me.