
Open source AI gateway routing every model behind one zero-markup endpoint.
Experiential is an open source AI gateway that gives developers, agents, and companies one OpenAI-compatible endpoint for every model they call. Hosted providers, bring-your-own provider keys, and local models all sit behind api.experientiallabs.ai/v1, routed at provider list price with no markup on routed tokens.
The gateway adds routing policy, provider failover, budgets and hard caps, key management, model allowlists, and per-agent token attribution, and it is self-hostable through the experiential package on PyPI. A paid intelligence layer on top of the same traffic suggests cheaper models per prompt, improves cache hit rates, and fine-tunes a model the customer owns on its own traces.
Source: experientiallabs.ai
Experiential Labs sells into a growing line of AI infrastructure spending: companies routing increasing model traffic across several providers at once and losing track of what each agent, team, and prompt costs. Its pitch is that none of that spend becomes an owned asset, and that a gateway plus a trained model converts it into one.
The buyer base the company describes is AI budget holders: platform teams with growing inference bills, startups on credit programs, and enterprises that want provider-independent routing, zero data retention routing, and a model they own. It prices hosted access by credit rather than by markup on tokens, and offers an enterprise tier with committed credits, SSO, RBAC, and private networking.
Source: experientiallabs.ai
Experiential's gateway publishes its source under Apache 2.0 and charges nothing on top of routed tokens, so a customer's model bill stays at provider list price while the routing, budgets, and attribution layer stays inspectable. That combination is unusual: closed commercial gateways charge a percentage of traffic, and raw provider accounts leave spend unattributed.
Because every request passes through one trace format, the same gateway that routes traffic also supplies the data its intelligence layer needs. The company reports a Rust data plane that adds about a millisecond per request, provider failover that keeps model identity fixed, and a path from routing to a fine-tuned model the customer owns rather than rents.
Source: experientiallabs.ai