Glonce

Experiential launches an open source gateway that turns usage into a custom router or model

toolsThe Glonce editorsHacker News🔥63Glonce rating63Impact45Novelty38Relevance92Surprise28Players22Credibility82
Experiential launches an open source gateway that turns usage into a custom router or model

A Hacker News user going by SilenN submitted a Show HN post for Experiential, a newly shown open source gateway and router for AI agent workflows. The project's own description gives it three jobs: put hosted, bring-your-own-key (BYOK), and locally run models behind a single OpenAI-compatible API; let an operator control which users and agents can call which models, for which use cases, and how much each can spend; and turn production traffic into a custom router or model optimized for quality, speed, and cost.

Running it locally starts with pip install experiential, then the exp command. On first run, a setup wizard walks through shared provider, model, and reasoning-effort selectors, persists every provider connection that gets configured, and then shows defaults for a public alias, an identity, and a $50.00 command budget before printing a one-time API key. The walkthrough then picks a public alias such as opus-5, exports the issued key as an environment variable named EXP_GATEWAY_KEY, and sends a bearer-token request to http://127.0.0.1:8000/v1/chat/completions, naming that alias as the model. The exp (or exp run) process itself is described as a compiled native data plane that serves every route on loopback, meaning it is reachable only from the local machine rather than the network.

For teams who would rather not run their own instance, a hosted alternative at platform.experientiallabs.ai exposes the same OpenAI-compatible API, plus the Anthropic Messages API, at api.experientiallabs.ai/v1. To lower the setup effort further, a file called SETUP.md is written specifically to be handed to a coding agent, such as Claude Code, Cursor, or Codex, as one of four copy-paste prompts, so the agent runs the whole setup itself rather than a person doing it by hand: creating an account and importing existing LLM traces as telemetry; connecting the user's own OpenAI, Anthropic, Gemini, Azure, Bedrock, Fireworks, or OpenRouter keys for free BYOK pass-through; making a first call through the gateway; and a fuller onboarding prompt that creates an account, connects keys, imports existing spend, and repoints every coding agent, including Claude Code, Cursor, Codex, and Aider, or Conductor, at the gateway.

The project's differentiator from a plain multi-provider proxy is what it does with recorded usage. A user first collects OpenTelemetry traces from their existing agent, or, to try the pipeline without their own data, downloads a public sample trace dataset published on Hugging Face under the experiential-labs namespace, named wmo-terminal-tasks-traces. The walkthrough then runs exp build support-agent, which steps through providers, models, and budget and fits a router simulation against those traces. A separate command, exp optimize model support-agent, goes further and fine-tunes an open source model the user owns, using a tool called Tinker. In Python, a fitted router loads back in as a standard OpenAI client instance backed by its own private gateway, for example with exp.load_router('my-project').

Anonymous, aggregate product telemetry, through PostHog, is enabled by default; the project states it never includes prompts, traces, actions, observations, paths, model names, credentials, or raw customer content, and it can be checked, disabled, or re-enabled with the commands exp config telemetry status, disable, or enable, with the preference stored locally in a file at .exp/settings.toml. No company, team, or funding behind Experiential is named in the text; only the platform.experientiallabs.ai and api.experientiallabs.ai domains, and the Hugging Face namespace experiential-labs, identify who built it. No license is named for the project despite the text calling it open source, and no release date or version number is given anywhere.

Key facts

  • Experiential is an open source gateway and router that puts hosted, BYOK, and local LLMs behind one OpenAI-compatible API, with controls over which users and agents can call which models, for which use cases, and how much they can spend.
  • A local install, pip install experiential followed by the exp command, runs a setup wizard that persists every selected provider connection and shows a default $50.00 command budget before issuing a one-time API key.
  • A hosted alternative at platform.experientiallabs.ai exposes the same OpenAI-compatible API, plus the Anthropic Messages API, at api.experientiallabs.ai/v1.
  • A file called SETUP.md hands a coding agent, such as Claude Code, Cursor, or Codex, four copy-paste prompts that walk it through creating an account, connecting BYOK provider keys for free pass-through, calling the gateway, and repointing existing coding agents at it.
  • Recorded OpenTelemetry traces can be turned into a custom, optimized router with the exp build command, or used to fine-tune an open source model the user owns with exp optimize model, using a tool called Tinker.

Why it matters

Experiential's own Hacker News launch title bills it as an open, self-hostable answer to OpenRouter. What moves it beyond a plain multi-provider proxy is a pipeline that closes the loop from usage back into the router itself: instead of a person picking a model for each request by hand, the project records production traces with OpenTelemetry, and two commands turn those traces into either an optimized router tuned for quality, speed, and cost, or a fine-tuned open source model the user owns. Paired with governance controls over which users and agents can call which models, for which use cases, and how much they can spend, this targets teams running many models across many agents that currently have no single place to see or control that traffic, and no automatic way to turn what that traffic shows into a better setup.

Who it affects

The audience is developers and teams operating LLM-backed agent workflows across more than one provider, especially anyone already juggling hosted, BYOK, and local models who wants a single OpenAI-compatible entry point with usage and spend controls attached. It is built to be set up largely by a coding agent rather than a person: SETUP.md hands the whole configuration to an agent such as Claude Code, Cursor, or Codex as copy-paste prompts, and the fullest of those prompts repoints existing coding agents, including Claude Code, Cursor, Codex, and Aider, or Conductor, at the gateway once it is running. Anyone who prefers not to depend on a hosted account can still run the whole thing locally, since the local gateway serves every route on loopback.

How to use it

There are two entry points. Locally, pip install experiential followed by the exp command starts an OpenAI-compatible gateway; on first run a setup wizard steps through provider, model, and reasoning-effort selectors, persists every configured provider connection, and shows defaults for a public alias, an identity, and a $50.00 command budget before printing a one-time API key, which the walkthrough then uses, with an example alias such as opus-5, to call the gateway at 127.0.0.1 on port 8000. Hosted, platform.experientiallabs.ai runs the same idea as a managed service, exposing an OpenAI-compatible API, plus the Anthropic Messages API, at api.experientiallabs.ai/v1. For the upgrade path from a plain gateway to a tuned router or model, collect OpenTelemetry traces from an existing agent, or pull the public sample dataset published on Hugging Face under experiential-labs; the build command steps through providers, models, and budget and fits a router against those traces, and a separate optimize command fine-tunes an open source model the user owns, using a tool called Tinker. No license is named anywhere for the project, despite the text calling it open source, so its exact reuse terms are not stated.

How solid is it

This account rests entirely on the project's own documentation, submitted to Hacker News as a Show HN post; every detail here, from the setup wizard's defaults to the telemetry exclusions, is the project describing its own design in its own words. No company, team, or funding behind Experiential is named in the text: only the platform.experientiallabs.ai and api.experientiallabs.ai domains, and the Hugging Face namespace experiential-labs, identify who is behind it. No release date or version number appears anywhere, and no adoption figures, such as user counts, downloads, or GitHub stars, appear in the text either.

Risks and caveats

The free BYOK pass-through for outside provider keys carries no stated cap on request volume or spending. Anonymous, aggregate PostHog telemetry is on by default, though opting out takes a single command, and the project states which categories of data it excludes from that telemetry. The router and fine-tuning pipeline is only as good as the traces behind it: both the build and the optimize steps depend on the user already having, or downloading, a working set of recorded OpenTelemetry traces before either command produces anything.