Agent frameworks

Switchyard

NVIDIA-NeMo/Switchyard

Route LLM traffic and translate OpenAI ↔ Anthropic payloads natively.

Repository
Stars
1.2k
Forks
85
Open issues
12
Language
Rust
Licence
Apache-2.0
Created
2026-07-01
Last push
2026-08-15
Effort to adopt

Some wiring

A few days of integration work — credentials, data mapping, a deploy target.

01

What it actually does

Switchyard is a blazing fast proxy server written in Rust that routes and translates LLM traffic. If you have an AI agent hardcoded to use OpenAI's API, but you want to run it against Anthropic's Claude or a local Llama 3 model, Switchyard sits in the middle and flawlessly translates the API payload. It handles streaming, tool calling, and structured outputs natively, allowing you to swap out closed-source models for open-source alternatives without rewriting your agent's code.

02

Who it's for

  • 01

    AI engineering teams who want to test their agents against different LLMs without refactoring code

  • 02

    Enterprises migrating away from OpenAI to local, open-source models for privacy reasons

  • 03

    Ops teams needing a central gateway to rate-limit, log, and load balance LLM API requests

EngineeringOperationsSaaSAgency
03

Where it earns its keep

  • Running Claude Code or Cursor against a locally hosted Llama 3 model
  • Creating a central LLM gateway for your entire engineering team to monitor API spend
  • Gracefully falling back to Azure OpenAI if your primary Anthropic API key hits a rate limit
04

Use it, or skip it

Reach for it when

  • You have an agent hardcoded to the OpenAI spec but want to use Claude
  • You need enterprise-grade telemetry and routing for your internal LLM usage

Skip it when

  • You are just writing a simple python script—just use LiteLLM or the native SDKs instead of running a proxy server
  • Your agent requires highly specific model features (like Gemini's massive context window) that don't translate well
05

10 automations

  1. 01Operations

    Cost-Optimized LLM Router

    Route simple summarization tasks to Llama 3, while routing complex coding queries to Claude 3.5 Sonnet, saving 80% on API costs.

  2. 02Engineering

    Automated API Fallback

    When OpenAI goes down, Switchyard automatically re-routes all traffic to Anthropic, ensuring zero downtime for your agents.

  3. 03Operations

    Centralized Audit Logger

    Log every single prompt and completion across your entire organization into BigQuery for compliance auditing.

  4. 04Engineering

    PII Redaction Proxy

    Agent strips credit cards and SSNs from prompts at the proxy layer before they ever reach an external API.

  5. 05Engineering

    Rate Limiter for Internal Agents

    Prevent runaway infinite loops in your agents by enforcing a strict budget cap per minute at the proxy level.

  6. 06Engineering

    Local Dev Override

    When engineers run agents locally, automatically route traffic to a cheap local model, but route to GPT-4 in production.

  7. 07Founders

    A/B Testing Framework

    Randomly split 50% of your customer support agent traffic to Claude and 50% to OpenAI to measure resolution times.

  8. 08Engineering

    Token Caching Layer

    Cache exact duplicate prompts at the edge so frequent identical queries (like CI/CD checks) return instantly and cost nothing.

  9. 09Engineering

    Semantic Routing

    Use a tiny local model to classify the prompt intent, then route code queries to DeepSeek and creative queries to Claude.

  10. 10Engineering

    Shadow Deployment

    Copy production LLM traffic and run it against a new open-source model asynchronously to evaluate its performance before switching.

Want one of these running by Friday?

LimeDock builds these as real workflows inside your stack — deployed to your cloud, wired into your Slack and CRM, with the code in your repo. You pay a build fee and your own API keys, nothing else.

Book a workflow call
07

Source

Repository stats were read from the GitHub API and reflect the last time we refreshed this entry. The editorial breakdown above is LimeDock’s own analysis — we are not affiliated with NVIDIA-NeMo.

https://github.com/NVIDIA-NeMo/Switchyard