RailRouter 🚂

A local LLM switchyard for macOS. One native app to see every model on your Mac (Ollama, LM Studio, MLX), and one OpenAI-compatible endpoint that routes each request to the right engine.

Français · Project page

What it does

• Library — every model from Ollama and LM Studio in one table: source, parameters, quantization, size, capabilities. Group, sort, search, copy a model id with one click.
• In memory — what is loaded right now (including direct MLX servers), with RAM use, activity state and, when the engine exposes it, prompt-processing progress.
• Discover — search the Ollama, Hugging Face (GGUF/MLX) and LM Studio catalogs, filter by capability, sort by popularity, newest or name, then download.
• Endpoint — a background service on http://127.0.0.1:41235/v1 that speaks the OpenAI API and routes by the model field:
  • Ollama (11434), LM Studio (1234), a running MLX server (8090)
  • Ollama cloud models, and free OpenRouter models (:free)
  • /v1/models, /v1/chat/completions, /v1/responses, /v1/embeddings, /health, /telemetry
  • /v1/systemone (Jev/SystemOne body: state + typed questions) and /v1/decisions (OpenAI chat body + choices, answered as a chat.completion with confidence/probabilities), model clef → the Clef MLX decision server (8091), which the gateway never starts

The endpoint listens on loopback only and keeps running after the window is closed (user LaunchAgent local.RailRouter.gateway).

Requirements

• macOS 14 or later, Swift 5.9+ toolchain (Xcode or Command Line Tools)
• At least one of: Ollama, LM Studio, an MLX server

Quick install

curl -fsSL https://raw.githubusercontent.com/Hugo291/RailRouter/main/install.sh | bash

Clones the repository to a temporary folder, builds, copies the app to /Applications and opens it. Needs the Swift toolchain (xcode-select --install).

Build and install

make test      # built-in test harness, no model is loaded
make install   # builds, signs ad-hoc, copies to /Applications
open /Applications/RailRouter.app

Using the endpoint

curl http://127.0.0.1:41235/v1/models
curl http://127.0.0.1:41235/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"model":"<an id from /v1/models>","messages":[{"role":"user","content":"Hello"}]}'

Point any OpenAI-compatible client (Codex, scripts, editors) at http://127.0.0.1:41235/v1.

OpenRouter free models need an OpenRouter key, read from OPENROUTER_API_KEY or the auth file used by opencode; the client's own Authorization header is never forwarded.

Limits worth knowing

• Counters come from real responses, never from an estimate of text length.
• Only LM Studio exposes a precise prompt-processing percentage; Ollama does not.
• Requests that bypass the endpoint cannot be tracked.
• Discover groups model families; it does not guarantee identical weights or quantizations across engines.
• The telemetry log keeps ids, phases, dates and counters only — no prompts, answers, cookies or tokens.

License

MIT