for teams shipping ai products
the right model, every call
map every llm call site, test models on your real prompts, and route each call to the cheapest one that provably holds quality — receipt included, no redeploy.
the workspace
see every call, together
simple enough to read at a glance. detailed enough to fix the routing.
The control plane
your app, every model — one layer in between.
the toolkit
everything you need to run ai in production
run models side by side, see what every call costs, and cap spend.
find the model that fits
fire the same prompt at every candidate model in parallel. outputs, latency, and cost land side by side — you pick what ships.
a log line for every call
every request through the gateway is logged — model, tokens, cost, latency, status. filter by workflow and see exactly where the spend goes.
change models without a deploy
point your sdk at the gateway once. pick the winner in the dashboard and every call with model: auto uses it on the next request.
budgets that actually stop spend
per-product caps enforced at the gateway. when the budget is gone, traffic stops — not just a dashboard alert.
integrations
fits the stack you already have
openai-compatible endpoint, mcp for your editor, and every model we serve behind one key.
start routing with proof
connect a repo, run your prompts across every model, and publish the routing that wins.
common questions
how cran drops into your product, in plain terms.
01What is Cran?+
A control plane for your product's AI: one gateway for every model, workflow discovery from your codebase, parallel model comparison on your real prompts, per-call logs, and routing you publish from a dashboard — no redeploy.
02How is it different from an ai gateway?+
Gateways route whatever you send them. Cran also knows your call sites, lets you run the same prompt across models in parallel, and logs every call's tokens and cost — so routing decisions are yours to make, with evidence.
03How hard is it to get started?+
Two commands: cran login, then cran mcp install for your editor. After that, saying "use cran" in any repo wires the connection, registers call sites, and sets up model routing.
04Does my source code leave my machine?+
No. Discovery runs locally over MCP. Cran stores workflow metadata — slug, prompt shape, call site — never your repo or customer payloads.
05Can I change models without redeploying?+
Yes. Publish a new model or prompt from the dashboard; calls with model: auto pick it up on the next request.