numinous-technology/open-ultra — explained in plain English
Analysis updated 2026-05-18
Reduce your Claude Code or Codex API bill by routing easy tasks to cheaper models automatically.
See a report of how often cheap models match your main model's quality before trusting them with real traffic.
Run coding agent traffic through a local proxy that keeps all data and trained routing on your own machine.
Test whether budget-tier models can handle your specific coding workload before switching over fully.
| numinous-technology/open-ultra | 0-bingwu-0/live-interpreter | 010zx00x1/faresnipe | |
|---|---|---|---|
| Stars | 2 | 2 | 2 |
| Language | Python | Python | Python |
| Setup difficulty | moderate | moderate | easy |
| Complexity | 3/5 | 2/5 | 2/5 |
| Audience | developer | general | general |
Figures from each repo's GitHub metadata at analysis time.
Requires a uv install, an API key for your frontier model, and running openultra init before routing can be trusted.
open-ultra sits between your coding assistant, such as Claude Code, and the AI model it talks to, with the goal of quietly switching some of your requests over to cheaper AI models once it has proven those cheaper models can match the quality of your usual, more expensive one. This is meant to save money on API costs without you having to notice a drop in quality. Here is how it works in practice. Every request from your coding agent still goes straight through to your main, expensive model exactly as before, so nothing about your day-to-day experience changes at first. At the same time, a hidden copy of each request is also sent to one or more cheaper models running in the background. A separate judge model then compares what the cheap model produced against what your main model actually did, either by comparing answers directly, or, for coding tasks, by letting the cheap model work in an isolated, throwaway copy of your project and comparing the resulting code changes. Once a cheap model has built up enough of a track record of matching the main model on a particular kind of task, you can turn on live routing, and the tool will start sending that type of request to the cheap model instead, only falling back to the expensive model for tasks the cheap one has not proven itself on yet. The tool is installed as a command-line program and works with several existing coding assistants, including Claude Code, Codex, and OpenCode, either by launching them through a wrapper command for a single session or by editing their settings so every future session is routed through the tool automatically. Setup involves an initial configuration step to add an API key and choose a starting set of cheap models, after which you can check reports on how often the cheap models are matching your main model and how much money that would save, before deciding whether to turn routing on. All the collected data and trained routing decisions stay on your own machine rather than being sent anywhere else. The project is licensed under the Apache 2.0 License.
open-ultra is a proxy that sits in front of your coding AI assistant and gradually shifts requests to cheaper AI models once they prove they match your main model's quality, cutting API costs.
Mainly Python. The stack also includes Python, uv, Anthropic API.
Use, modify, and distribute freely, including commercially, as long as you keep the license notice and attribution.
Setup difficulty is rated moderate, with roughly 30min to a first successful run.
Mainly developer.
This repo across BitVibe Labs
Don't trust strangers blindly. Verify against the repo.