Flux OS — AI skill for Claude Code

"hi"                                          → gpt-oss-20b      (groq)     ~$0.00004
"Translate 'good morning' into Spanish"       → mistral-small-4  (mistral)  ~$0.0002
"Write a Python function that parses dates"   → gpt-oss-20b      (groq)     ~$0.0004
"Implement a lock-free concurrent hash map in
 Rust, thread-safe, no global locks, ..."     → o4-mini          (openai)   ~$0.0053

Most traffic doesn't need your most expensive model. flux-os reads each request (task type, difficulty, tools, images, JSON mode, context size), sets a quality bar, and picks the **cheapest model that clears it**. If that model fails (rate limit, outage, timeout), the request is **rerouted** to the next-best model, on a different provider first.

• Zero code changes. Point your OpenAI or Anthropic client at flux-os and set model="auto".
• Agents work. Tool calls, streaming tool calls, and multi-turn tool loops work across every provider, whichever API format your framework speaks.
• Three API formats in, any provider out. Chat Completions, the Responses API and Anthropic Messages, translated to OpenAI, Anthropic, Gemini, Groq, Mistral, DeepSeek, OpenRouter, Ollama or any OpenAI-compatible server.
• Fast. Pure-Python heuristics, no LLM in the decision path, well under 1 ms per decision.

Quickstart

pip install "git+https://github.com/vbc1406/flux-os"

# See what it would do. No API keys