[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$f1qtubu6jw78vy":3},{"code":4,"message":5,"data":6,"timestamp":34},10000,"success",{"full_name":7,"position":8,"description":9,"language":10,"stars":11,"forks":12,"stars_gained":13,"age_days":14,"hn_threads":15,"one_liner":22,"summary_md":23,"topic":24,"github_topics":27,"license":28,"homepage":29,"updated_at":30,"history":31},"experientiallabs\u002Fexperiential",11,"Experiential is the open source, zero markup gateway for BYOK, self-hosted and 1000+ marketplace models. It learns from your traffic to cut costs, recommend better models, and train a specialized model you own.","Python",12831,263,1224,123,[16],{"id":17,"title":18,"points":19,"comments":20,"posted_at":21},49471407,"Show HN: We built open OpenRouter that turns usage into a better model",222,47,"2026-08-27","A self-hostable gateway that routes agent traffic across providers, then turns collected traces into custom routers or models.","If you're already routing agent traffic and want the traces to become a reason to standardize on one gateway, this is worth an afternoon. If you only need a model proxy, simpler options exist.\n\n## The problem\n\nAn agent team starts by calling one provider directly, then adds keys for a second and a third as quality and cost trade-offs appear. Very quickly nobody can say which users, agents or features are burning which budget, and swapping a model requires editing each app. Experiential puts one OpenAI-compatible endpoint in front of hosted, bring-your-own-key and local providers, with identity and spend control at the gateway. The request path stays the same while model choice moves out of application code.\n\n## What's different\n\nThe distinguishing step is what happens after routing: the gateway collects OpenTelemetry traces and offers an `exp build` flow that turns them into a simulated project and a fitted router, plus an `exp optimize model` path for fine-tuning an open model via Tinker. That converts production traffic into a tuning input rather than leaving it as dashboards only. A public sample dataset of terminal-task traces is provided, so you can try the build step before you have your own traffic. The trade-off is that recommendation and optimization depend on having representative traffic; a fresh install with a handful of requests has little to learn from, and following the tuning path sends your traces to a third-party fine-tuning service.\n\nSetup is deliberately flexible, which cuts both ways. `pip install experiential` and `exp` starts a local loopback gateway with a setup wizard, a default alias and a generated key; the hosted platform serves the same API if you'd rather not run it yourself. The [SETUP.md prompts](https:\u002F\u002Fgithub.com\u002Fexperientiallabs\u002Fexperiential\u002Fblob\u002Fmain\u002FSETUP.md) hand the whole configuration to a coding agent. The local install is easy to try; keeping a self-hosted gateway upgraded and its provider connections healthy is ongoing work.\n\nThere is also a separate `exp capture` mode that records traces from the Codex and Claude Code desktop apps using your existing subscriptions. The README labels it experimental and notes macOS network reliability is still under investigation, so it is not the part to build a workflow around.\n\n## Who should try it\n\nGood fit: a team running agents across several providers that wants central budget and identity control, and is prepared to feed traces into routing experiments afterward. A solo developer who mostly wants a local proxy and a single key should start with the local `exp` path and ignore the optimization features. Poor fit: someone who needs capture to be dependable today, since the README itself calls it experimental and macOS-limited, or a team unwilling to export traces outside its own infrastructure for tuning.\n\n## Requirements and limitations\n\nThe documented local path needs Python and installs a CLI; the hosted route needs an account and an `xpl_` key. Capture is the most constrained piece: macOS, Python 3.13+, and approval for a browser extension and certificate. Anonymous aggregate telemetry is on by default and must be disabled explicitly if that matters to you.\n\nTwo things to watch. The optimization workflow asks for trace files, so its value scales with how much real traffic you can share, and the fine-tuning step depends on Tinker as an external service. The README excerpts cover routing, capture and tuning but not the policy details of per-user spend enforcement or failure behavior when a provider is down; those are worth reading in the [project documentation](https:\u002F\u002Fgithub.com\u002Fexperientiallabs\u002Fexperiential\u002Fblob\u002Fmain\u002FREADME.md) before putting it in front of real users.\n\n## HotRepo's take\n\nThe repository was created in late June 2026 and reached 12,389 stars by 2026-10-25, the single daily-board observation HotRepo has, when it added 1,224 stars and ranked 12th on the new-repository board. That is one snapshot of a young project, not a trend line, so persistence cannot be judged yet. The [Show HN thread](https:\u002F\u002Fnews.ycombinator.com\u002Fitem?id=49471407) drew 222 points and 47 comments; one commenter noted the usage dashboard screenshot appeared to be hosted-only rather than in the repo [on 2026-08-28](https:\u002F\u002Fnews.ycombinator.com\u002Fitem?id=49483508:), which is a reasonable thing to check if self-hosting is your goal. Verify whether the router-fitting workflow can run entirely against your own trace exports.",{"slug":25,"name":26},"ai-engineering","AI Engineering",[],"Apache-2.0","https:\u002F\u002Fexperientiallabs.ai","2026-10-25T11:01:30Z",[32],{"date":33,"stars":11,"stars_gained":13},"2026-10-25",1792971680]