Run Claude Code, Cursor, and Codex on any model via one proxy

What it does

Lynkr runs a local HTTP server that impersonates the Anthropic API. Point Claude Code, Cursor, or the Codex CLI at localhost:8081 and the proxy forwards requests to a backend of your choice—whether that’s a free local Ollama instance, AWS Bedrock, Databricks, or OpenRouter. The client tools need no code changes; you override a base URL and supply a dummy API key.

The interesting bit

The tier-routing feature classifies each request as simple, medium, complex, or reasoning, then dispatches it to a different model. The README claims this pushes 70–90% of traffic toward cheaper or local models while reserving expensive ones for genuinely hard tasks. It is a pragmatic admission that not every autocomplete needs a frontier model.

Key highlights

• Supports 12+ providers, including 4 local options (Ollama, llama.cpp, LM Studio) for fully offline use
• Drop-in compatibility: override ANTHROPIC_BASE_URL for Claude Code, or set a model provider override in Cursor and Codex configs
• Built-in token optimizations (prompt caching, semantic caching) and load shedding under memory pressure
• 699 tests passing, Apache 2.0 licensed
• Enterprise-friendly routing to Databricks, Azure OpenAI, and AWS Bedrock

Caveats

• Getting Claude Code and Cursor to play nice requires explicit compatibility shims: you must cap POLICY_MAX_STEPS and POLICY_MAX_TOOL_CALLS and disable POLICY_SAFE_COMMANDS_ENABLED, suggesting the proxy has to massage Anthropic-specific behavior that other models don’t natively support
• The README dedicates an entire section to common startup errors, including missing tier warnings, fallback credential complaints, and a pino-pretty transport bug in older versions—so expect some initial friction

Verdict

Worth a look if your team loves Claude Code’s UX but needs to run local models, hit a company-mandated Bedrock or Databricks endpoint, or slash API spend. Skip it if you’re already content with Anthropic’s native pricing and performance.