CowAgent Review 2026: Self-Evolving Assistant, Ex chatgpt-on-wechat

What CowAgent is and who it's for

CowAgent is an open-source AI assistant and agent harness. The repository at github.com/zhayujie/CowAgent describes it as a "super AI assistant" that plans tasks, runs tools and skills, and self-evolves with memory. It was formerly known as chatgpt-on-wechat, and that history matters more than the marketing copy suggests.

The name change is the story. chatgpt-on-wechat started as a bridge: put an LLM behind a WeChat account and let people chat with it. CowAgent keeps that bridge but adds a task loop on top. The assistant now plans, calls tools, remembers things across sessions, and can be extended with skills. It is still open source, still self-hostable, and the pricing page lists a starting tier of $0/mo because the software itself costs nothing — you pay only for the model API calls you route through it.

Who is this for? Three groups, roughly.

First, developers and tinkerers who want a personal assistant reachable from the chat apps they already live in. If your team runs on Feishu or DingTalk, or your personal life runs on WeChat and Telegram, CowAgent meets you there instead of asking you to install another app.

Second, people who want an agent they can inspect and modify. Closed assistants hide the loop. CowAgent's loop is in the repository. You can read how it plans, what it sends to the model, and where memory gets stored. For anyone who has been burned by a black-box assistant quietly changing behavior, that transparency is the point.

Third, small teams that want a shared assistant without per-seat SaaS pricing. Because you self-host, the cost model is your infrastructure plus API usage, not a subscription that scales with headcount.

Who is it not for? People who want zero setup. CowAgent is a self-hosted project. The docs describe configuration files, API keys, and a running process. If you want to sign up and start typing in ninety seconds, this is the wrong tool. That is not a flaw — it is the trade you make for control.

From chatgpt-on-wechat to a general agent harness

The original chatgpt-on-wechat had a narrow job. It connected a chat platform to a model and relayed messages. Useful, popular, and limited. Every new capability — image generation, voice, web search — arrived as a bolt-on.

CowAgent reorganizes around a harness. The docs describe a loop where the assistant receives a request, plans a sequence of steps, selects tools or skills to execute those steps, observes results, and iterates until the task is done. That is a different architecture from a message relay. A relay answers one prompt. A harness pursues a goal.

The practical difference shows up in multi-step work. Ask a relay "what's on my calendar tomorrow and should I bring an umbrella" and you get a model guessing. Ask a harness and it can call a calendar tool, call a weather tool, then reason over both results. The harness is what makes tool use coherent instead of decorative.

The rename also signals ambition beyond WeChat. "chatgpt-on-wechat" locked the project to one platform and one model family in the public imagination. "CowAgent" is platform-neutral and model-neutral. The repository keeps the old name in its history and in the migration path, which is the honest way to do a rename: don't pretend the past didn't happen, just make the new thing better.

One caveat worth stating plainly. Agent harnesses are harder to debug than relays. When a relay fails, the message didn't send. When a harness fails, it might have planned wrong, picked the wrong tool, misread a tool's output, or looped. The docs cover configuration and skills, but anyone running this in production should expect to spend time reading logs. That is true of every harness in this category, not a CowAgent-specific defect.

Channels: WeChat, Telegram, Feishu, DingTalk, web

Channel support is CowAgent's inheritance and its advantage. The repository and docs list WeChat, Telegram, Feishu, DingTalk, and a web interface among the supported surfaces.

WeChat is the deepest integration because it is the original one. For anyone in a WeChat-centric environment — much of China's consumer and small-business world — this is close to unique. Most Western agent frameworks treat WeChat as an afterthought or skip it. CowAgent treats it as a first-class channel, which is a direct consequence of its lineage.

Telegram covers the opposite audience: developers and international users who treat Telegram as their default messaging layer. Bots are straightforward there, and the channel fits the self-hosted model well.

Feishu and DingTalk push the project into workplace territory. These are the collaboration platforms where Chinese companies actually operate. An assistant that lives in Feishu can be pulled into work threads, which is where a task-running agent earns its keep. A chat assistant that only answers trivia is a toy; one that can be summoned inside a work channel and asked to run a skill is closer to infrastructure.

The web interface matters for a less obvious reason: it is the escape hatch. When a platform's API changes, rate-limits you, or bans your bot account, the web surface still works. Having a channel that no third party controls is a resilience feature, not a checkbox.

What the docs do not promise is parity. Channel capabilities differ. Rich media, group behavior, and message formatting vary by platform, and anyone planning a deployment should read the per-channel notes rather than assume one configuration works everywhere. The honest framing: CowAgent gives you many doors into the same assistant, but each door has its own lock.

Skills, tools, and long-term memory

Skills and tools are how a harness does real work. The docs describe an extensible system where capabilities are added as skills the agent can invoke. This is the standard pattern in modern agent design, and CowAgent follows it: the model decides what to do, the harness provides the verbs.

The interesting part is memory. CowAgent's description says it "self-evolves with memory." In practice, long-term memory means the assistant retains information across sessions instead of starting fresh each conversation. That changes the interaction from transactional to cumulative. You tell it once that you prefer short answers, or that a certain project lives in a certain repo, and it carries that forward.

Memory is also where agent projects most often disappoint. Storage, retrieval, and relevance are hard. A memory system that recalls everything is as useless as one that recalls nothing, because the context window fills with noise. The docs describe the memory feature; they do not claim it is perfect, and no honest review should either. What I can say from the documentation is that the capability exists and is central to the project's identity, not a footnote.

Skills plus memory plus a planning loop is the combination that separates an agent from a chatbot. A chatbot with memory is a chatbot that remembers. An agent with memory is something that accumulates competence at your specific tasks. That is the promise CowAgent is making, and the architecture described in the repository is consistent with it.

For builders, the extension story is the practical question. If adding a skill means writing a small module and registering it, the project is genuinely extensible. If it means patching core files, it is not. The docs describe a skills system, which points to the former, but anyone with a specific integration in mind should read the skill-authoring documentation before committing.

Model support and self-hosting

CowAgent is model-agnostic by design, which is the correct choice for a harness. The model is a component you swap, not the product itself. The docs describe support for multiple providers, and because you supply your own API keys, the cost of running CowAgent is the cost of the models you choose.

That choice matters more than it used to. The live pricing snapshot shows the spread clearly. At the top, openai/o1-pro lists at $150 per million input tokens and $600 per million output tokens, with a batch tier at $75 in and $300 out. A step down, anthropic/claude-opus-4.7-fast and openai/gpt-5.5-pro both list at $30 in, with outputs at $150 and $180 respectively. openai/gpt-5.2-pro sits at $21 in and $168 out. openai/o3-pro lists at $20 in and $80 out. openai/gpt-5-pro is $15 in and $120 out. anthropic/claude-opus-4.1 and claude-opus-4 both list at $15 in and $75 out. openai/o1 is $15 in and $60 out. And older models like openai/gpt-4 and openai/gpt-4-0314 sit at $30 in and $60 out.

Why does this matter for a CowAgent review? Because a harness multiplies token usage. A single user request can trigger several model calls as the agent plans, acts, and observes. If you wire a $600-per-million-output model into an agent loop and let it run unattended, the bill compounds fast. The sensible configuration for most people is a mid-tier model for the planning loop and a cheaper one for routine steps, or a batch tier where the workload allows it. openai/gpt-5.5-pro:batch lists at $15 in and $90 out, which is a meaningful discount over its standard tier for work that tolerates latency.

Self-hosting is the other half. Because CowAgent runs on your infrastructure, your conversation data and memory store stay where you put them. For individuals that is a privacy nicety. For companies with data-residency rules, it is often the deciding factor. The trade is operational: you own uptime, upgrades, and backups. The docs describe deployment; they do not run it for you.

One honest note on cost expectations. The software is free — the pricing page lists $0/mo as the starting tier — but "free" here means no license fee, not no cost. API usage is real money, and self-hosting has an infrastructure cost even if it is small. Anyone budgeting should model token spend first, because that is the line item that scales with use.

GitHub stars, repo health, release cadence

The repository at github.com/zhayujie/CowAgent shows 47,126 stars. That number is the single strongest signal in this review, and it deserves context rather than applause.

Roughly 47,000 stars puts CowAgent in the top tier of open-source AI assistant projects. Stars accumulated partly under the chatgpt-on-wechat name, which was one of the earliest and most visible ways to put an LLM behind a chat account. Popularity that predates a rename is still popularity — the users came for a reason and many stayed.

What stars do not tell you: whether the project is maintained, whether issues get answered, whether releases ship on a predictable cadence. A dormant repo can carry 47,000 stars for years. So the star count is a starting point, not a verdict. Anyone evaluating CowAgent for real use should look at the commit history, the issue tracker, and the release tags directly. The repository is public; the evidence is there to check.

What I can say from the repository signals is that the project is active enough to have executed a full rename and repositioning, which is not something abandoned projects do. A rename touches documentation, package names, configuration, and user expectations. Doing it at all implies maintainers who are still investing.

The size of the community cuts both ways. A large user base means more skills, more answers in issues, and more eyes on bugs. It also means more surface area, more legacy configuration, and a higher chance that a specific question has been asked and answered in a thread you have to find. For a self-hosted project, community size is a real asset — it is the closest thing to support you get without paying for it.

My read: 47,126 stars is a strong signal of relevance and a weak signal of current quality. Treat it as a reason to look, not a reason to trust. The docs and the repository are where the actual evaluation happens.

Verdict: who should use CowAgent and who shouldn't

CowAgent is a credible, well-positioned open-source agent harness with an unusual advantage: a genuine foothold in WeChat, a platform most competitors ignore. The 47,126 stars are earned, the architecture described in the docs is sound, and the $0/mo starting tier means the only real cost is the models you route through it.

Use it if you want an assistant inside the chat platforms you already use, especially WeChat, Feishu, or DingTalk. Use it if you want to read and modify the agent loop rather than trust a black box. Use it if you are comfortable self-hosting and would rather pay for tokens than seats. Use it if long-term memory and a skills system are features you will actually exercise.

Skip it if you want zero setup. Skip it if you need vendor support with an SLA. Skip it if you are unwilling to monitor token spend, because an agent loop with a premium model attached can get expensive quickly. And skip it if you need a guarantee of channel parity across every platform, because the docs describe per-channel differences rather than uniform behavior.

For teams already comparing self-hosted assistant options, Beetlix is our own product, and it approaches the same problem from a different angle — worth a look if you want a managed path alongside a self-hosted one. But if control and platform reach are your priorities, CowAgent is one of the more interesting open-source options in 2026, and the rename from chatgpt-on-wechat reflects a real shift in what the project is trying to be.

How this review was researched

This review draws on the CowAgent repository at github.com/zhayujie/CowAgent, the vendor documentation and site at cowagent.ai, the official pricing page, and the live model pricing data referenced above. No hands-on testing was performed; claims about behavior come from the project's own documentation and repository signals, and are attributed as such.