Cursor thinks it's talking to OpenAI. It's not.

What it does

cursor-deepseek is a local Go proxy that intercepts OpenAI-shaped API calls from Cursor IDE and reformats them for DeepSeek, OpenRouter, or Ollama. It handles streaming responses, function calling, and message format conversion, then pipes the translated replies back to Cursor’s Composer. The README notes it runs on port 9000 and expects a publicly reachable endpoint—via ngrok, Cloudflare Tunnel, or similar—so Cursor’s cloud service can reach your local instance.

The interesting bit

Cursor is apparently locked to OpenAI’s API schema, and this project treats that as a protocol problem rather than vendor lock-in. The proxy is essentially a dedicated adapter: separate Docker builds and binaries for each provider, with hardcoded model aliases like gpt-4o quietly remapped to DeepSeek equivalents.

Key highlights

• HTTP/2 with Brotli, Gzip, and Deflate compression
• Translates streaming and tool-calling payloads between OpenAI and provider-specific formats
• Server-side API key validation and CORS headers
• Docker images split by provider variant (DeepSeek, OpenRouter, Ollama)
• GPLv2 licensed

Caveats

• Only /v1/chat/completions and /v1/models are supported; other OpenAI endpoints are absent
• Model mapping is vague—the README says gpt-4o maps to “DeepSeek’s GPT-4o equivalent model” without naming the actual target model
• You must choose one provider at build or run time; there is no apparent runtime switching between DeepSeek, OpenRouter, and Ollama

Verdict

Worth a look if you already pay for Cursor Pro and want to route Composer to a cheaper or self-hosted backend. Skip it if you need a general-purpose OpenAI gateway or multi-provider fallback logic.