Tarmac is a policy-aware routing layer for AI agents. Set the perimeter, hand your agents a budget, and we'll find them the cheapest fit across every provider — without you babysitting the bill.
One endpoint. Tarmac picks the provider that meets your latency, region, and quality bar at the lowest price — and re-picks the moment a cheaper seat opens up.
Each agent gets an identity, a budget, and a policy. Spend caps, model allowlists, region locks, working hours — set once at the gate and enforce them everywhere.
Tarmac watches every request and quietly downshifts to a smaller model or shorter context when the answer doesn't need the big iron. Same output, a fraction of the bill.
One environment variable. Drop-in compatible with the OpenAI, Anthropic, and OpenRouter SDKs. No wrapper library to maintain.
TARMAC_URL=https://api.gettarmac.dev/v1
Identity, budget, allowed models, allowed regions, working hours. YAML or dashboard — whichever your team prefers.
agent: research-7b · cap: $40/day
Every request is priced against every provider in real time. Tarmac routes to the seat that meets the spec at the lowest price.
avg routing decision · 11ms
Per-agent spend, model mix, drift from cap, retried requests, held requests — all on one dashboard, all in one invoice.
live audit log · per-agent breakdown
Sign in to the control tower. Issue keys, edit policies, audit per-agent spend, and watch routing decisions in real time.