Experiential: open-source LLM gateway that turns your traces into a custom router

Experiential is an open-source, self-hostable gateway that sits in front of large language models and exposes a single OpenAI-compatible endpoint (it also speaks the Anthropic Messages API). Teams can route requests across hosted, bring-your-own-key, and local models, and enforce policy on top: which users and agents may call which models for which use cases, plus per-key spending caps. A pip install experiential and the exp command spin up a loopback gateway with a setup wizard, a default command budget, and a one-time API key; a managed version is offered for those who would rather not run it themselves.

The pitch that distinguishes it from a plain proxy like OpenRouter is the feedback loop. Experiential ingests OpenTelemetry traces from your existing agents, builds a simulation from that production traffic, and optimizes a project-specific router against it for quality, speed, and cost. From there it can fine-tune an open-source model you own via Tinker, so real usage data becomes a tailored router or model rather than just billing telemetry. Fitted routers load directly as an OpenAI client backed by a private gateway.

BYOK support spans OpenAI, Anthropic, Gemini, Azure, Bedrock, Fireworks, and OpenRouter, with free pass-through, and the docs include copy-paste prompts that let coding agents such as Claude Code, Cursor, and Codex run the setup and repoint themselves. Anonymous aggregate product telemetry via PostHog is on by default but, per the project, excludes prompts, traces, model names, credentials, and any raw customer content, and can be disabled with a single command. The practical appeal is consolidating access control, budgeting, and routing for agent fleets in one place while keeping the door open to owning the resulting model.