OpenTeam groups five web AI models in one chat
An article describes OpenTeam, a self-reported Chrome extension that brings the web versions of ChatGPT, Claude, Gemini, DeepSeek and Grok into a shared group chat. Members can @ one another, see each other’s replies and challenge one another. The extension uses logged-in web sessions rather than API keys, with per-site adapters, isolated iframes and DOM-based completion detection because the sites do not expose a finish_reason signal. It assigns context budgets separately for each model and supports independent and collaborative modes. The article also documents limitations: reliance on DOM selectors, exposure to site risk controls, no token-level usage data, no tool execution, limited support for Chromium-based browsers, and a PolyForm Noncommercial 1.0.0 license that permits personal, non-commercial use. Its capabilities and implementation details have not been independently verified.
The article provides a concrete engineering account of coordinating multiple AI web sessions without APIs, including site-specific adapters, model-specific context budgets and a three-stage reply-completion heuristic. It also makes clear that this approach trades API reliability and control for reuse of existing logged-in subscriptions.
上周做一次技术选型,同一个问题我问了五个模型,拿到五份答案。最后我采纳的,是我一开始就想选的那个。
这件事比"模型答错了"更麻烦:单模型真正的问题不是它会错,而是你没有办法发现它错。 你只有一个视角的时候,那个视角就是全部现实。你只能靠自己的判断力去校验它,而你的判断力恰好是这整件事里最弱的一环。
于是我做了个东西:把 ChatGPT、Claude、Gemini、DeepSeek、Grok 的网页版拉进同一个群聊,@ 谁谁发言,多个 AI 能看到彼此说过的话,可以互相反驳。
它叫 OpenTeam,一个 Chrome 扩展。这篇文章不讲"我做了个",讲三个工程问题:为什么不用 API、群聊是怎么跑起来的、以及最难的那个——你怎么知道一个网页 AI 已经回答完了。
一、为什么不用 API
先说我放弃了什么。