The latest articles on DEV Community by Wetidom (@wetidom). https://dev.to/wetidom https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4055169%2Feca852ab-c202-480d-92a2-7ffce5b2682d.jpg https://dev.to/wetidom en Wetidom Thu, 30 Jul 2026 13:30:04 +0000 https://dev.to/wetidom/how-i-built-a-multi-provider-proxy-for-grok-build-4kg2 https://dev.to/wetidom/how-i-built-a-multi-provider-proxy-for-grok-build-4kg2 <p>I <strong>wanted</strong> to use Ollama and other providers with Grok Build without <br> being locked to one API. So I built GrokRoute — a local proxy that <br> routes requests to 5+ providers with auto-fallback.</p> <p><strong>How it work</strong>s</p> <p>Grok Build → GrokRoute (:8083) → Agnes / Groq / Zhipu / Laguna / Ollama</p> <p>The proxy is a single Python file implementing the OpenAI <br> /chat/completions endpoint with SSE streaming. When a request comes in:</p> <ol> <li>Proxy reads the provider config</li> <li>Tries providers in priority order</li> <li>If one fails (rate limit, timeout) — next provider takes over in 5s</li> <li>Streams the response back to Grok Build</li> </ol> <p>Non-streaming mode goes even faster: fires requests to ALL providers <br> in parallel, first valid response wins.</p> <p><strong>Quick start</strong></p> <p>npm install -g grokroute<br> grokroute install<br> grokroute</p> <p>Then press SPACE — Grok Build launches with multi-provider support.</p> <p><strong>Key features</strong></p> <ul> <li> <strong>5+ providers</strong>: Ollama, Groq, Agnes, Zhipu, Laguna</li> <li> <strong>Auto-fallback</strong>: sequential with 5s timeout per provider</li> <li> <strong>Race mode</strong>: parallel requests, fastest response wins</li> <li> <strong>Web admin UI</strong>: dark theme, :8083/admin</li> <li> <strong>Local-only</strong>: runs on 127.0.0.1, no telemetry</li> <li><strong>MIT license</strong></li> </ul> <h2> Why I built this </h2> <p>I was benchmarking different models for coding tasks and got tired of <br> switching configs manually. The auto-fallback turned out to be the <br> killer feature — if a provider rate-limits, it just works.</p> <p>Source: <a href="https://github.com/wxstdo-boop/grokroute" rel="noopener noreferrer">github.com/wxstdo-boop/grokroute</a></p> <p>Happy to answer questions!</p> ai opensource python showdev