AI Bridge

AI Bridge turns a tab you are already logged into on an AI chat website into a local, OpenAI-compatible endpoint for a local IDE or agent. It deliberately keeps the two halves on your computer:

IDE / agent ──HTTP──> 127.0.0.1:4317/v1 ──WebSocket──> Chrome extension ──DOM──> logged-in AI chat tab

Supported AI chats

• Grok
• Gemini
• DeepSeek
• Perplexity
• Qwen
• ChatGPT
• Claude
• Z.ai
• Kimi

This project borrows the practical idea of a web-chat-to-OpenAI gateway from geminiweb-as-api and the extension-first direction of Quarkonix, without copying browser profiles or exposing a remote server.

Install

1. Install Node.js 20 or newer.

2. From this directory, start the local companion with Run.bat or manually in cmd:

   npm start

   On first start it creates .ai-bridge-token next to local-server.mjs, then prints that bridge token. Leave this terminal running. Future starts reuse the same token, so the extension does not need to be configured again.

3. In Chrome, open chrome://extensions, turn on Developer mode, choose Load unpacked, and select the extension folder in this project.

4. Open the AI Bridge extension, paste the printed token, and click Save & connect.

5. Click a provider in the popup, sign in normally, and leave that tab open. Choose it as the default provider, or use a provider-specific model name.

The server binds only to 127.0.0.1. It never reads, copies, or uploads Chrome profile files. Each request is serialised because a chat tab can answer only one prompt at a time. While connected, the extension exchanges a heartbeat every 20 seconds, so its Manifest V3 service worker remains active indefinitely on Chrome 116 or newer. The Pause button intentionally stops that connection; Continue restores it.

After changing extension files during development, click the extension's Reload button on chrome://extensions. Restart npm start after changing the local server.

Use from an IDE

Set the IDE's OpenAI-compatible base URL to:

http://127.0.0.1:4317/v1

Use bridge-auto to route to the configured default provider, or select one explicitly:

bridge-grok       bridge-gemini     bridge-deepseek
bridge-perplexity bridge-qwen       bridge-chatgpt
bridge-claude     bridge-zai        bridge-kimi

The local endpoint implements GET /v1/models, POST /v1/chat/completions, including Server-Sent Events when stream: true, and GET /health.

Automatic “name this conversation” requests are detected and completed locally; they do not use the active AI-chat tab. Set BRIDGE_INTERCEPT_TITLES=0 before starting the server to disable this behavior.

curl http://127.0.0.1:4317/v1/chat/completions `
  -H 'Content-Type: application/json' `
  -d '{"model":"bridge-gemini","messages":[{"role":"user","content":"Explain this error briefly."}]}'

Most OpenAI clients require a non-empty API key even when a local server does not. Supply any placeholder value unless you opt into local API-key protection:

$env:BRIDGE_API_KEY = 'choose-a-local-secret'
npm start

With BRIDGE_API_KEY set, send Authorization: Bearer choose-a-local-secret.

Bridge-token storage

The bridge token is automatically stored in the ignored .ai-bridge-token file. Delete that file to rotate the token, then paste the newly printed value into the extension. To use a managed token without writing a file, set BRIDGE_TOKEN; to choose another token-file location, pass --token-file <path> or set BRIDGE_TOKEN_FILE.

Settings

ai-bridge.settings.json sits beside the local server and is deliberately readable and versionable:

{
  "interceptTitleRequests": true,
  "injectPromptInstructionsEveryMessage": true,
  "prePrompt": "Generic Pre-prompt",
  "postPrompt": "Generic Post-prompt",
  "// disguiseModelNameAsAssistant": "Turning on may increase responsiveness of a model to the prompt",
  "disguiseModelNameAsAssistant": false
}

Set interceptTitleRequests to false to pass IDE conversation-naming calls through to the selected AI website. The environment variable BRIDGE_INTERCEPT_TITLES=0 overrides this setting for one run. Use BRIDGE_SETTINGS_FILE or --settings-file <path> to load a settings file from elsewhere.

Set disguiseModelNameAsAssistant to true to always replace {{modelName}} with an assistant instead of the provider display name. Turning this on may increase responsiveness of a model to the prompt. Keys that start with // are comments and are ignored.

Prompt instructions

Markdown files under prompt-instructions/ are wrapped around the bridged chat transcript before it is typed into the AI tab:

prompt-instructions/
  pre-prompt/
    Generic Pre-prompt.md
    ZCode Pre-prompt.md
  post-prompt/
    Generic Post-prompt.md
    ZCode Post-prompt.md

• prePrompt / postPrompt name the active file stem (with or without .md). Set either to "" to disable that side.
• Pre-prompt text is placed at the very start of the outgoing prompt; post-prompt text at the very end.
• With injectPromptInstructionsEveryMessage: true (default), both are injected on every chat completion. Set it to false to inject only on the first user turn of a conversation (messages has at most one role: "user" entry).
• Use {{modelName}} in either file for the provider display name (for example Claude, Gemini, ChatGPT). Prefer a provider-specific model (bridge-claude, …) so the server can substitute immediately; with bridge-auto, the placeholder is filled by the content script from the live tab hostname.
• Edit the markdown files anytime; the next request reloads them. Settings changes still require restarting npm start.
• Use BRIDGE_PROMPT_INSTRUCTIONS_DIR or --prompt-instructions-dir <path> to point at another instructions root.

Notes and limits

• This uses the visible web UI, not a provider's private or undocumented network API. UI changes can require selector updates in extension/content.js.
• You are responsible for the terms, plan limits, and acceptable-use rules of every provider you choose to connect.
• File uploads, images, provider-specific citations, and native tool execution are outside the first version. When the IDE sends a tools array, the bridge best-effort detects JSON tool-call text from the chat tab (bare, fenced, or prose-wrapped) and returns OpenAI tool_calls for both streaming and non-streaming requests.
• Claude responses are intentionally emitted once its final response block is stable, rather than streaming its intermediate thinking/status blocks into the IDE.
• Keep the requested provider page fully loaded. If a request says the content script is not ready, reload that tab.
• The extension has permissions only for the nine named chat websites and localhost; it does not have broad <all_urls> access.

Development

npm test

The tests cover the local health endpoint, model list, title-request interception, bridge relay, streaming, heartbeats, token persistence, prompt-instruction injection, and text-to-tool_calls conversion.