One URL in front. The right machine behind it.
Every request goes whole to the machine that already has its model loaded and room to run it. Nodes find the router on your network by themselves.
Run Ollama across multiple computers: Macs, Linux servers and Windows PCs, NVIDIA or not, as one endpoint for Claude Code, Codex and Open WebUI. Free and open source.
See what the herd runs · See the biggest herds
Install it and your fleet joins the leaderboard within a day.
Every request goes whole to the machine that already has its model loaded and room to run it. Nodes find the router on your network by themselves.
Two environment variables put Claude Code on the fleet. One provider block does it for Codex. Any model id auto-routes to the best model you have loaded, with no model map to write.
Claude Code guide →LLMs, embeddings, image generation, speech-to-text and vision. Each request goes to a machine that can serve it, so transcription lands on an Apple Silicon node running Qwen3-ASR and image jobs on a node with an image backend.
Multimodal routing →Routing works the same everywhere, NVIDIA or not. Apple Silicon adds MLX, image generation and speech-to-text.
Model loaded, memory fit, queue depth, latency, session affinity and more. Your follow-up returns to the warm cache.
With capacity learning on, it steps off a Mac when you join a call, and away from a machine busy with a build or a game.
OpenAI, Ollama and Anthropic APIs on one URL: Open WebUI, LangChain, Aider, Continue, Codex.
Stop leaving compute on the table. Start herding.