Turning Cursor’s web chat into an API—when it stays up

What it does

This Go service sits between your OpenAI-compatible client and Cursor’s web interface, translating chat/completions requests into whatever Cursor’s frontend expects and piping back the responses. It exposes gemini-3-flash and automatically mints a gemini-3-flash-thinking variant, plus supports tools, tool_choice, and streaming. The whole thing is written in Go but carries a Node.js runtime to execute JavaScript code—likely for browser fingerprinting or token generation.

The interesting bit

The README spends as much space on failure modes as on features, which tells you everything about the architecture. There is explicit glue logic for agent orchestrators like Kilo Code that throw a fit if a model sees tools but doesn’t call them: enable KILO_TOOL_STRICT and the service will retry non-streaming requests once to force a tool invocation.

Key highlights

• Presents Cursor’s web backend as a standard OpenAI chat/completions endpoint.
• Auto-derives *-thinking models from any base model you configure.
• Supports function calling with tools, tool_choice, and tool_calls.
• Ships with a non-commercial PolyForm license—commercial use is explicitly banned.
• Bundles a Node.js dependency to execute JavaScript, suggesting the auth flow relies on frontend-style obfuscation.

Caveats

• The repository description warns “这次倒下啦” (it fell this time), suggesting the upstream Cursor web interface may have already blocked or broken the integration.
• Troubleshooting docs openly list 403 errors, Cloudflare interception, and token acquisition failures as routine hazards.
• Requires Node.js alongside Go, which is an unusual runtime pairing for a simple API proxy.

Verdict

Worth a look if you want to experiment with free Gemini Flash access through standard OpenAI clients and don’t mind maintaining a fragile scraper. Everyone else—especially anyone seeking a stable, production-grade API—should keep their credit card handy.