Most stacks leave the model boundary open: hand-picked models, budgets as dashboards you read after the money is gone, provider lock-in. AgenticGateway closes exactly that. Agents call one OpenAI-compatible endpoint with one key; tenants, budgets, upstream credentials and routing all derive from it server-side. Routing aliases resolve against a policy built from an AgenticPerformance eval run — change the eval run and you change the routing, with the eval run recorded for audit. Cost is a circuit breaker, not a report: tenant-window and per-run ceilings deny fail-closed the moment they trip, and every call lands in a ledger and as a cost-attributed OpenTelemetry span. Transport — failover, load-balancing, retries, streaming, semantic caching — is Bifrost's job, with microsecond overhead; adding a provider is config, not code.
The two surfaces most agent stacks leave open — model & provider and cost & FinOps — implemented properly, and nothing else.
Agents call one OpenAI-compatible endpoint with a single gateway key; tenants, budgets, upstream credentials and routing derive from it server-side, and client keys are stored as sha256 only.
Route aliases resolve against a policy built from an AgenticPerformance eval run — the ranking becomes the failover chain, the eval run id is recorded, and no LLM sits on the hot path.
Tenant-window and per-run ceilings are enforced in code and deny the moment they trip — cost control as a circuit breaker, not a monthly report.
Provider abstraction, failover, load-balancing, retries, streaming and semantic caching are handled by Bifrost with microsecond overhead; adding a provider is configuration, not code.
Every call lands in a hash-not-text ledger and as a cost-attributed OTel span in AgenticPerformance's conventions — cost-per-outcome in the same traces as your evals.
AgenticMind's guard runs inline, protected-path tool calls are checked against AgenticAssurance's policy pack, and a sampled judge reviews high-risk routes.
To help you with any questions that are not listed here, we offer free, no-obligation consultations
A proxy forwards calls. AgenticGateway is a policy plane: routing chosen from measured eval runs, cost ceilings enforced fail-closed in code, guardrails on the boundary and an auditable evidence trail — the Standard's Model & provider and Cost & FinOps surfaces, implemented.
Everything transport: provider abstraction, failover, load-balancing, retries, streaming and semantic caching, with microsecond overhead. AgenticGateway never re-implements transport — adding a provider is config, not code.
Apache 2.0, on GitHub (Moai-Team-LLC/AgenticGateway). Free to use, self-host and build on, including commercially.
Routing policies are built from AgenticPerformance eval runs, and every call is emitted as a cost-attributed OTel span in APL's conventions — so routing decisions and their outcomes live in the same measured loop.
It's a Bun-native TypeScript service on a Bifrost data plane. Your agents don't change — they keep speaking the OpenAI-compatible API, just against one endpoint with one key.
More from our agentic practice — the standard we build to, the open-source layers that implement it, and the team that ships it.
Running a fleet against multiple model providers? Tell us about it and we'll show you what one governed key can do.