Momentum you can audit.
Real milestones with real dates. No marketing quarters, no "coming soon" entries — everything below shipped and is verifiable from the API surface.
What shipped, when.
2026-10-09
Decision models — the Jev protocol, governed
Jev/Laya-style decision models are a first-class model type on the canonical POST /v1/systemone wire protocol: state + typed questions in, calibrated answers out — all three Jev primitives (choice with per-option probabilities + confidence, score on an ordered rubric, noul as a raw 0..1 yes-probability). Decision traffic flows through the SAME pipeline as LLMs: alias-group load balancing and failover across Jev providers (hosted TypeSafe, self-hosted laya-serve, any compatible backend), budgets, per-token or per-question billing, and one HMAC-signed event_type='decision' row per call — a blocked decision is never invisible. Upstream Jev usage (input_tokens/output_tokens) maps into canonical token accounting so hosted-Jev bills real counts.
Also shipped: the decisions playground — a request-builder with ready-to-run Jev workflow sets (support-ticket triage, frustration & urgency, issue routing, citation checks, and the full 22-question guard matrix), threshold verdicts computed live and re-verified server-side; and runnable decision clients in examples/decisions/.
2026-08-09
CEL routing rules + complexity router + adaptive LB
Rule-based routing shipped: CEL expressions over request context (model, complexity_tier, budget_used, headers, params) with ordered targets and chain rules, a configurable complexity classifier, and adaptive load balancing with retry budgets — all wired into the routing pipeline and manageable from the dashboard's Routing Rules page.
2026-08-08
Zero-content stream failover
A stream that dies before any content frame — clean EOF without [DONE], or an upstream read failure — now fails over transparently on the same client SSE stream. The cutting provider is marked unhealthy despite its HTTP 200. Also shipped: the capture-provenance stamp on every event, the playground provider-routing control, and system-call visibility for title/suggestion generation.
2026-08-07
Alias-group load balancing + provider directives
Attach one alias to models on multiple providers and Trifrost load-balances across the group: rr_health, weighted, or least_latency, with a 30s cooldown and half-open recovery. The provider request field gained per-request directives — only / ignore / order — matched by slug and stripped before upstream.
2026-08-04
Request metrics
Time-to-first-token and tokens-per-second captured end-to-end on every inference event, plus an hourly provider-performance rollup — compare providers on measured latency and throughput, not brochures.
2026-07-14
Enterprise governance
Per-model budget matrix in USD and tokens across Customer → Department → User → Virtual Key. Model allowlists with intersection semantics and explicit deny-all. Structured 402 responses. Zero database hits on the governance hot path.
2026-06-30
Billing & spend
Token accounting with dual cost tracking, pricing snapshots per event, Holt forecasting, spend alerts, and a six-tab spend dashboard — by provider, model, department, user, key, or session.
2026-06-26
Governance hierarchy + dashboard
The Customer → Team → Virtual Key hierarchy with budgets, rate limits, and RBAC. HMAC-signed audit logs. OIDC provisioning. A full dashboard UI uplift to operate all of it without touching a config file.
2026-06-24
Plugin pipeline foundation
The priority-ordered PreLLM/PostLLM hook pipeline with short-circuit semantics that everything since has been built on. Plus the provider and model catalog — 27 provider kinds with a capability matrix and pricing sync — and virtual keys (sk-bf-*), hashed at rest with timing-safe comparison.
Verify it in the API
Every milestone above is observable from the surface: failover in attempt_trail, metrics in event rows, budgets in structured 402s.
Read the docs →See it in the console
Playground, spend, events, providers, governance — everything shipped since June is one click away.
Launch console →Comparing us?
The category difference between a router and a control plane, in eight facts and one table.
Trifrost vs OpenRouter →Shipping now
Universal gateway · alias-group load balancing + zero-content failover · governance hierarchy with per-model budgets · model allowlists · HMAC-signed audit log · OIDC/SSO with claim-driven teams · playground · session cost attribution · spend forecast + alerts · five regions.
On the public roadmap
Guardrails (secrets + PII, external providers) · prompts repository · skills repository · MCP gateway · OTLP export · HA / clustering.
Tell us what your enterprise needs. Watch it land here.
Deployed in AU · APAC · EMEA · UK · DE. Need somewhere else? We'll stand it up.