Hi HN,

I built an open-source proxy called Lynkr that lets AI coding tools like Cursor, Cline, Continue.dev, Claude Code, and OpenCode run on any model backend instead of being locked to a single provider.

The problem I ran into is that most modern AI coding tools are tightly coupled to a small set of hosted models. That makes it hard to control cost, run locally, use enterprise endpoints, or experiment with different models without changing tools or workflows.

Lynkr sits between the client and the model and speaks both OpenAI-compatible and Claude-compatible APIs. From the client’s point of view, nothing changes — but you can route requests to local models (Ollama, llama.cpp), OpenRouter, or enterprise endpoints like Azure or Databricks.

Architecture-wise, it’s a thin HTTP proxy that translates requests, supports streaming, tool calls, and adds optional features like prompt caching, simple routing rules, and workspace indexing for coding workflows.

One feature I’ve found especially useful is hierarchical routing: requests can be routed through a priority chain of models (for example, try a local model first, fall back to a stronger cloud model on failure or for large contexts, and finally to a premium model for complex tasks). This makes it possible to mix cost, latency, and quality policies without changing the client.

A few concrete use cases that motivated this:

Running Cursor or Continue fully on local models for privacy

Routing cheap tasks to local models and harder tasks to stronger cloud models

Using Claude Code CLI without sending code to Anthropic’s hosted backend

This is early-stage and very much infra-focused. Setup is a single binary / npm install, and you point your tool’s base URL at Lynkr.

Repo: https://github.com/Fast-Editor/Lynkr