Building a Real-time API Proxy for Cross-Platform LLM Tooling Interoperability

The Challenge: Bridging Disparate AI Ecosystems

The core problem addressed is the lack of interoperability between proprietary AI tool ecosystems. Cursor IDE, while feature-rich, is designed to route all its AI agent requests through its own billing and backend infrastructure, specifically tailored for Anthropic models. This creates a 'walled garden' scenario, preventing users from leveraging existing subscriptions like GitHub Copilot with Cursor's advanced UI features. The solution involves creating a Man-in-the-Middle (MITM) proxy to intercept, modify, and redirect traffic, effectively making Cursor communicate with a different LLM provider than it expects.

Proxy Architecture Overview

The proxy architecture consists of two main local services working in tandem to achieve the desired re-routing and payload transformation. This distributed setup ensures separation of concerns: one service handles authentication and acts as a bridge, while the other intercepts and modifies network requests.

1. Port 4141 (The Auth Layer / Copilot Bridge): An open-source connector responsible for handling the OAuth handshake with GitHub Copilot and exposing an OpenAI-compatible endpoint. This acts as the translation layer for authentication and API compatibility.
2. Port 4142 (The Intelligence Proxy / Router): A high-performance proxy server (implemented with Bun) that intercepts Cursor's outbound HTTP traffic. Its primary roles are to manipulate request payloads and forward them to the Copilot Bridge (Port 4141).

Bypassing Internal Routing Tables

A significant architectural challenge was Cursor's internal routing logic, which hardcodes known model names (e.g., 'claude') to its proprietary backend, overriding user-defined custom API endpoints. The solution involved discovering a "loophole": prepending a unique prefix (e.g., `cus-`) to the model name. This tricks Cursor into not recognizing the model, causing it to gracefully fall back to the user-specified local proxy URL. The proxy then strips this prefix before forwarding the request, ensuring the target LLM receives the correct model identifier.

let json = await req.json();
const PREFIX = "cus-";

// The Heist: Strip the prefix so Copilot gets the real model name
if (json.model && json.model.startsWith(PREFIX)) {
  const targetModel = json.model.slice(PREFIX.length);
  console.log(`🔄 Rewriting model: ${json.model} -> ${targetModel}`);
  json.model = targetModel;
}

Real-Time AST Mutation for Schema Compatibility

The most complex part of the system design was handling the differences in API schema for tool calling. Cursor's frontend sends Anthropic-flavored tool schemas, which are incompatible with GitHub Copilot's strict OpenAI-flavored function call expectations. This required a recursive Abstract Syntax Tree (AST) cleaner within the proxy to intercept, traverse, and mutate the JSON payload in real-time. The cleaner removes illegal JSON schema properties (like `additionalProperties`, `$schema`, `title`) and translates the schema structure (e.g., `input_schema` to `parameters` within a `function` object) to ensure Copilot's API accepts the request without errors.