Experiential launched on Hacker News as an open-source OpenRouter that feeds usage logs back into model selection and routing logic. The project sits at experientiallabs/experiential and already drew 157 points and 26 comments.
What It Is
Experiential mirrors the OpenRouter API surface while logging every prompt, response, and latency metric. It then runs lightweight preference modeling on those logs to adjust routing weights toward models that performed better on similar tasks. No central company controls the collected data.
How It Works
Users point their applications at the local Experiential endpoint instead of the hosted OpenRouter. Each completed request stores prompt embeddings, chosen model, token counts, and user feedback signals. A background job retrains a small routing head every few hours, shifting traffic toward higher-performing models for that usage cluster.
Benchmarks and Specs
The repository currently reports no public latency or accuracy numbers. Early testers on the Hacker News thread noted that the routing updates require at least several thousand requests before measurable shifts appear. Memory footprint stays under 4 GB when running the default routing service alongside a local inference server.
How to Try It
Clone the repository and run the provided Docker compose file. Point existing OpenAI-compatible clients to http://localhost:3000/v1. The README contains a one-line curl command that starts both the router and a sample preference trainer.
Pros and Cons
- Pros: Full data ownership, no per-token fees, routing adapts to private usage patterns.
- Cons: Requires enough local traffic to train the router; no pre-built model catalog like the hosted OpenRouter service.
Alternatives and Comparisons
| Feature | Experiential | OpenRouter (hosted) | LiteLLM Router |
|---|---|---|---|
| Open source | Yes | No | Yes |
| Usage-based routing | Yes | Yes | Basic only |
| Data stays local | Yes | No | Yes |
| Hosted option | No | Yes | No |
Who Should Use This
Teams already running multiple local or self-hosted models and collecting their own telemetry will benefit most. Solo developers or teams with fewer than a few hundred daily requests should continue using the hosted OpenRouter until they accumulate sufficient data.
Bottom Line / Verdict
Experiential gives organizations a practical way to close the loop between usage and routing without sending prompts to a third party.
The project remains early, but the core loop—log, retrain, reroute—addresses a clear gap between fully hosted routers and static local proxies.
Top comments (0)