Not another AI gateway
Your AI traffic isn't short on gateways. It's short on decisions.
Fluxen proxies OpenAI, Gemini, and Ollama traffic too — but that's the substrate, not the product. The product is finding what's inefficient, proving the fix against your real history, applying it safely, and confirming it actually worked.
one command, self-hosted
git clone https://github.com/ManojVihari/Fluxen.git
cd Fluxen && docker compose up --buildWondering how this is different from LiteLLM, Portkey, or a FinOps dashboard? See the comparison.
The core loop
1. Understand
Every request through the gateway is accounted for — cost, tokens, model, provider, latency, cache status. No sampling, no estimation.
2. Identify
Four detectors continuously look for real, evidence-backed inefficiency: model cost, repeated requests, token drift, and traffic anomalies.
3. Simulate
Before anything changes, Fluxen replays the scenario against your real historical traffic and shows the actual cost delta — not a guess.
4. Control
Apply a proven change with one human-confirmed action: percentage-based model routing, exact caching, budgets, rate limits, or model restrictions.
5. Measure
Every applied change gets a before/after verdict against real traffic 7 and 14 days later — successful, partial, no effect, or regressed, with one-click revert.
What makes this different
Evidence before action
Every opportunity shows its evidence — sample size, eligible traffic, confidence — before you're asked to do anything.
Simulation before production
Nothing changes production traffic until you've seen the real, replayed impact against your own history.
Measured, not assumed
Fluxen tells you what actually happened after a change, not just what it estimated beforehand.
Three providers, one gateway
OpenAI, Google Gemini, and Ollama — behind one OpenAI-compatible endpoint. Point your existing client at Fluxen and route, cache, and control traffic across all three without touching application code.