OpenTeam Puts Five AI Web Apps in One Group Chat to Surface What a Single Model Hides
Multi-model review no longer needs API keys or per-token spend when the web versions are orchestrated through DOM automation, but the cost is living without finish_reason, usage data, or DOM stability. Disagreement between five models exposes a boundary condition one model misses, and four models agreeing on the same error points at training-data bias instead of a single bad sample.
OpenTeam is a Chrome extension built as a group chat over the web UIs of five major AI assistants: ChatGPT, Claude, Gemini, DeepSeek, and Grok. Users @-mention any role, and every model sees the others' replies so they can challenge each other. The whole thing runs on subscriptions people already pay for; there is no API key, no per-token meter, and no SDK to rotate. The author's starting point: with a single model, an incorrect answer is undetectable because the only perspective you have is the model's, and your own judgment is the weakest check in the loop.
Engineering-wise, OpenTeam is a Manifest V3 extension that loads each AI site in a separate iframe (overriding X-Frame-Options and CSP through declarativeNetRequest), injects content scripts with all_frames: true, and implements a per-site ChatSiteAdapter for input, send, reply container, and completion. Because there is no finish_reason, completion is inferred from DOM text length stability, with tiered debounce thresholds, a separate rule for short replies, container repaint detection, and timeout compensation. Context is quarantined per role with character budgets that mirror each model's window (1M for Gemini and DeepSeek down to 256k for Grok), and roles first run independently, then collaboratively, to avoid an echo chamber.
The author is explicit about the trade-offs: DOM automation breaks when sites redesign, anti-bot systems may flag accounts, and large-scale parallel calls are impossible. The project is open source under PolyForm Noncommercial 1.0.0, supports only Chromium, and requires logged-in sessions. Installation is a three-step load-unpacked flow from GitHub Releases.
The author's opening anecdote is the core motivation: a single model's error is undetectable because the user only has that one perspective, and their own judgment is the weakest validator in the loop.
A flat subscription makes marginal token cost zero, which is why DOM automation is economically rational even though it is permanently fragile; the architecture trades stability for a pricing model that fits how people already pay for AI.
Short replies are the hardest case for completion detection because a few characters don't change between polls, so a naive debounce misclassifies them; the fix, a longer stability window, is a reminder that simple heuristics need edge-case-aware thresholds.
Running independent mode before collaborative mode preserves first reactions and prevents echo chambers; going collaborative first only produces five echoes of the loudest opinion, which is a danger any multi-agent orchestration should design around.
Four models converging on the same wrong answer is more informative than one model's unique catch because it suggests the error lives in training data rather than in a single sampling; that is a useful signal for anyone evaluating models.