Experiential

Experiential is an open source AI gateway built by Experiential Labs, a Y Combinator-backed company, that lets teams access every major AI model through a single OpenAI-compatible endpoint. The gateway is written in Python (with a compiled native data plane), licensed under Apache 2.0, and available both as a self-hosted deployment and a managed hosted service at api.experientiallabs.ai/v1.

What It Is

Experiential sits in front of hosted providers (OpenAI, Anthropic, Google, Meta, Mistral, DeepSeek, and others), bring-your-own-key (BYOK) connections, and self-hosted local models, routing every request through one /v1 endpoint at the provider's list price with 0% token markup. On top of the routing layer sits an intelligence layer that watches traffic patterns and surfaces model-switching recommendations, caching opportunities, and — for high-volume workloads — a fine-tuned model trained on the customer's own traffic and served behind the same endpoint.

How the Gateway Works

The core routing model is straightforward: change the base URL to api.experientiallabs.ai/v1, keep existing OpenAI SDK code, and every model answers from one key. Key capabilities include:

• Provider failover — automatic fallback across providers and pooled accounts on capacity errors
• Hard budget caps — per-key spending limits enforced on every request, by day, week, or month
• Model allowlists — restrict which models a given key, agent, or team can call
• Zero Data Retention routing — policy-enforced routing only to providers under a ZDR agreement
• Full attribution — every token logged by agent, person, model, and day with live request logs showing route, cost, and TTFT

The gateway can be started locally with pip install experiential && exp, or pointed at the hosted platform with no infrastructure to manage.

The Intelligence Layer

Above the gateway, Experiential's intelligence layer adds three optimization paths described on the product site:

1. Model suggestions — traffic monitoring surfaces when a model switch would improve quality, speed, or cost, including new models on the day they ship
2. Caching — the system identifies cache-hit opportunities and shows projected savings; repeated tokens return at 90% off when enabled
3. Custom model ownership — for workloads that repeat at volume, Experiential can fine-tune an open source model on the customer's own traffic, validate it in closed-loop simulation, and serve it behind the same endpoint

The company publishes case studies on its website documenting measured results from this fine-tuning approach, including a computer-use agent and a claims-research workload.

Open Source Deployment Model

The gateway core is published at github.com/experientiallabs/experiential under the Apache License 2.0. The repository had approximately 4,485 stars and 165 forks as of the latest GitHub metadata. Self-hosting runs the full OpenAI-compatible /v1 endpoint, key management, budgets, provider waterfalls, and the usage API. The uvx --from experiential exp run command starts it on any infrastructure. Telemetry is anonymous aggregate data via PostHog and can be disabled with exp config telemetry disable.

Update: v0.7.67

The latest release as of the GitHub metadata is v0.7.67, published on 2026-09-11. The repository was created in June 2026 and has seen active development, with the last push also on 2026-09-11. The project direction signals continued investment in the intelligence layer (model optimization, caching, custom fine-tuning) alongside the open source gateway core.