WebModel Bridge

Universal local AI proxy turning 9 web AI platforms into free, standard OpenAI & Anthropic APIs with zero token costs and zero API keys.

Comment

💡 Inspiration

Developers, students, and hobbyists often want to test agents, IDE tools (like Cursor, Windsurf, or Claude Code), and automated workflows with top-tier AI models. However, official API access requires paid tokens, credit cards, and strict rate limits, even though these same models are completely free to use via their consumer web interfaces.

Existing reverse-engineered projects frequently break because they try to spoof API endpoints and get blocked by Cloudflare, WAFs, and bot detection. We built WebModel Bridge to solve this sustainably: automate a genuine Chrome browser session via the Chrome DevTools Protocol (CDP), inheriting authentic browser fingerprints, cookies, and TLS signatures to turn free web AI into standard developer APIs.

🚀 What It Does

WebModel Bridge is a local proxy daemon that bridges 9 major web AI platforms into standardized OpenAI (/v1/chat/completions) and Anthropic (/v1/messages) HTTP APIs with Server-Sent Events (SSE) streaming:

  • 9 Flagship Providers Supported: Anthropic Claude, OpenAI ChatGPT, DeepSeek, Google Gemini, xAI Grok, Perplexity, Moonshot Kimi, Alibaba Qwen, and Zhipu GLM.
  • Thinking / Scratchpad Filtering: Automatically detects and suppresses internal reasoning tags (<think>, status wrappers), streaming only finalized clean answers.
  • Interactive Command Center & Chat Studio: Built-in web dashboard at http://localhost:3456 with live session health, multi-provider chat, and a 1-click API Code Generator (cURL, Python SDK, Node.js, Claude CLI, Cursor).
  • Anti-Bot & Pacing Resilience: In-page error detection intercepts upstream quotas and rate limits, with adaptive typing emulation if modern web editors block clipboard pasting.
  • Origin-Partitioned Concurrency: Tab-level page pooling prevents cross-provider race conditions.

🛠️ How We Built It

  • Core HTTP & Streaming Engine: Built with TypeScript and Hono on Node.js 20+ for ultra-low latency SSE streaming and dual OpenAI / Anthropic format compliance.
  • Browser Automation Layer: Uses Playwright-Core connected directly to genuine Chrome instances via the Chrome DevTools Protocol (CDP).
  • DOM Event Observers: Custom MutationObserver injections track dynamic element generation, text node updates, and stream chunks in real time across 9 different web app frameworks.
  • Interactive UI: Responsive Dashboard and Chat Studio built with modern CSS and vanilla JS, with zero build-step overhead for instant local load times.
  • Testing & Diagnostics: Fully covered with 180 automated unit, integration, and E2E diagnostic tests using Vitest.

🧗 Challenges We Ran Into

  1. Diverse Web UI Structures: Every platform renders text differently—some stream raw text, others replace Markdown DOM nodes on every chunk, and some (like DeepSeek or Qwen) wrap responses in thinking blocks. We developed adaptive selector pipelines and token-level streaming normalizers.
  2. Clipboard Paste Blocking: Several platforms intercept paste events and require simulated human keystroke pacing (12ms delays).
  3. Session State Isolation: Managing concurrent browser tabs without cookie collisions required building an isolated origin page pool.

🏆 Accomplishments That We're Proud Of

  • Unifying 9 flagship AI platforms into single, drop-in replacement API endpoints.
  • Making modern developer tools like Claude Code CLI and Cursor run seamlessly against free web accounts.
  • Achieving 180/180 passing tests with 100% strict TypeScript types.
  • Zero credentials stored, zero telemetry, and 100% localhost-first privacy.

📚 What We Learned

We mastered deep Chrome DevTools Protocol internals, real-time DOM mutation streaming across disparate reactive web frameworks (React, Vue, custom Canvas/WASM), and building resilient fallbacks against anti-bot friction.

🔮 What's Next for WebModel Bridge

  • Multi-account round-robin load balancing.
  • Extended multimodal capabilities (voice input and audio generation).
  • Docker and headless server deployment modes.

Built With

Share this project:

Updates