Model connections in BrokenCode

Select BrokenGPT in the connection panel and complete authorization in the browser. BrokenCode stores the resulting credential locally. The BrokenGPT connection uses the stable broken-one model alias. Choose your connected account in the model picker before starting work.

1. Open Settings, then Model connections.
2. Choose Connect BrokenGPT.
3. Authorize the device in your browser and return to BrokenCode.
4. Confirm the connection indicator is green before starting a session.

In the desktop composer, open the reasoning control next to the model. Drag the slider or choose a named level. Options come from the selected model; Default keeps that model's normal behavior. On iPhone and Android, open Models & providers and use Reasoning.

Ultra is a coordinated team mode, not another provider reasoning level. It runs through the existing task runner and can use more tokens. Website connections use their website's default model and reasoning instead of these API controls.

This optional desktop feature is off by default. It connects a chatbot website to BrokenCode's existing task runner. It does not turn a website subscription into an official API. Your conversation, tool definitions, and tool results are sent to the service you choose; that service's account limits and terms apply.

• Built-in recognition covers ChatGPT, Claude, Gemini, Grok, Perplexity, DeepSeek, and Copilot. Recognition is not proof of compatibility. Websites, accounts, and locales can behave differently; failed checks stay out of the model picker.
• Add another chatbot by entering its public HTTPS home or new-chat URL. Automatic detection checks its message box, Send control and completed response. Ambiguous or unsupported layouts show a connection error; universal website compatibility is not guaranteed.
• Connections accept text only and use the website's default model settings. Images, PDFs, and audio require a suitable API model. Oversized turns fail explicitly instead of silently dropping context.
• Website replies can request tools; BrokenCode's normal task runner validates and executes them with your existing permissions. No separate browser agent loop is created.
• Sign-ins stay in separate local browser profiles. Login has no automation channel; verified tasks use a private CDP pipe with no exposed debugging port. Your personal browser profile is not imported. Turning the experiment off cancels pending work and removes its models. Restarting the desktop requires verification again.

1. Open Settings → Web chatbots and enable the experiment.
2. Paste the chatbot website, such as chatgpt.com or claude.ai, and choose Connect. Popular services also have one-click shortcuts. BrokenCode detects the chat controls automatically; no selectors or connection IDs are needed.
3. If needed, sign in in the separate Chrome, Edge, Brave or Chromium browser and choose Continue. Quitting that connection browser also starts verification automatically. Two small response checks use no project files. Saved sessions are checked when you reconnect.
4. Select the verified connection in the model picker. It also becomes available to the native phone apps connected to this computer.

• Never paste a token into brokencode.json or commit it to a repository.
• Disconnect from Settings to revoke the local BrokenCode link.
• Reconnect if authorization was interrupted or the local credential expired.

OpenCode remains first in the provider picker, and its free model catalog remains visible. Your explicit or saved model choice takes precedence. A free catalog listing does not guarantee inference access: OpenCode currently rejects some free-tier requests from BrokenCode. Keep your preferred model to retry later, or choose another connected model.

• If OpenCode reports that its free tier can only be used within OpenCode, open /models and choose another connected model.
• You can connect an OpenCode Zen API key for models available to your account.
• OpenCode Zen is a separate model connection selected by you. BrokenGPT continues to use the stable broken-one alias and private routing.