Route between models, ground them in your own corpus, gate every release on evals, and ship behind a canary.
Route, ground, gate and ship share one config and one trace, so every release carries its own evidence.
one line · four stagesOne endpoint in front of every model you use, with per-route budgets and automatic failover when a provider returns 429.
p95 overhead 41msAnswers come from your own corpus with citations that resolve, or they do not leave the run.
re-embed on writeEvals run on the diff, not on a schedule, and block the deploy when any threshold is missed.
240 cases · 3m 40sCanary a percentage, watch the same numbers in production, and roll back with one command.
rollback 1.2s medianMost teams bolt a router, a vector store, an eval harness and a deploy script together and spend the next year maintaining the seams. These four share one config, one trace and one set of numbers.
One endpoint in front of every model you use. Per-route budgets, automatic failover on 429 and 529, and a blended cost that updates as you change the mix.
Your corpus, chunked and embedded on ingest, queried with a floor you set. Every answer carries citations that resolve, or the answer does not leave the run.
An eval suite that runs on the diff, not on a schedule. Faithfulness, refusal rate, latency and cost, each with a threshold that blocks the deploy when it is missed.
Canary a percentage, watch the same four numbers in production, promote or roll back in one command. The rollback is a config change, so it takes seconds.
Seats are free. You pay for routed tokens and eval minutes, and the first of each every month is on us.
For the prototype that lives on one laptop and one API key.
For a feature in production that somebody is on call for.
For everyone whose agents outnumber their engineers.
No. The router takes the request you already send and returns what the provider returns; the only thing that changes is the base URL. Prompts, tools and streaming all pass through untouched, which is also what makes failover to a second model a config line rather than a rewrite.
The deploy stops and the previous config keeps serving. Nothing pages until a human asks it to. The failing cases, their scores and the diff that caused them are waiting in the run when someone looks.
On Bench and Line, in an encrypted store in the region you pick, and you can export the chunks and their embeddings at any time. On Foundry it never leaves your network — the collector and the index run inside your VPC and we never see the text.
No. Relay is a fictional product built as a design demo: the platform, the models, the numbers, the prices and the terminal output were all invented for it. Nothing is routed, nothing is stored, and no request leaves this page.
Put the four stages on one line and get your evenings back. Two minutes to the first routed call.