AI Briefing
KO

Switchyard: A Rust proxy for running Claude Code with open-source models

NVIDIA-NeMo/Switchyard

·2026.08.20 19:59

This is a Rust-based proxy that switches coding agents like Claude Code and Codex to open-source models such as vLLM, NVIDIA NIM, and Ollama. Clients send requests maintaining the existing OpenAI or Anthropic API formats, and Switchyard converts them into formats understood by the backend models for delivery. Responses are also translated back into the format expected by the client.

It supports intelligent routing rather than simple forwarding. Traffic can be distributed or used for A/B benchmarking through various algorithms, including LLM classifiers, signal-based tier routers, escalation routers, and random distribution. Operational metrics such as requests, errors, latency, and token usage can be tracked in real time via Prometheus metrics.

Two modes are provided: running as a standalone server or embedding as a Rust library. The library mode does not have its own HTTP stack and only exposes routing algorithms, making it easy to integrate into existing gateways or agent runtimes. Currently in the pre-alpha stage, APIs and algorithms may change significantly before v1.0, and production use is not recommended.

It is suitable for developers who want to flexibly distribute traffic across multiple models for cost and performance optimization. It is particularly useful for teams looking to apply open-source models to coding workflows without being dependent on closed-source APIs. It is released under the Apache 2.0 license.

GitHub
GitHub repository

NVIDIA-NeMo/Switchyard

Switchyard lets LLM applications route traffic across models and providers while preserving native OpenAI and Anthropic API compatibility - enabling flexible model selection, benchmarking, and cost/performance optimization.

Python

This introduction was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report errors, attribution issues, or removal requests via Contact.