Five endpoints that turn text, code and logs into calibrated probabilities your code can threshold, log and test — one model call per decision.
/v1/ask callPut fast, typed AI decisions on either side of your LLM, in CI, and in your alert pipeline — with thresholds you can read in every response.
Catch prompt injection, harmful requests, leaked secrets and off-topic messages going into — and coming out of — your LLM.
Send each prompt to the cheapest model tier that can handle it. Escalate automatically when confidence is low.
Check every diff against conventions written in plain English. Fail CI only on the rules you mark as errors.
Turn stack traces into a root cause, a severity, an owning team, and whether a retry will fix it.
Bring your own yes/no, choice and score questions over any text or JSON — the raw building block.
Each endpoint is a fixed set of narrow questions plus a policy in code — so you know exactly what you're getting before you call it.
{ "decision": "block",
"triggered": ["prompt_injection"] }
{ "tier": "small", "confidence": 0.86,
"escalated": false }
{ "passed": false,
"summary": { "fail": 1, "review": 1, "pass": 6 } }
{ "category": "database",
"severity_label": "Degraded" }
{ "answers": { "urgent": { "noul": 0.93 } } }
{ "jev_calls": 302, "input_tokens": 164512,
"by_endpoint": { ... } }
The model answers narrow questions. Your policy — a few thresholds in code — turns the answers into actions.
Every guardrail response shows which checks fired and how strongly. Switch between strict, balanced and permissive policies — or tune the thresholds — without changing a single question.
More on guardrailsChoice and score answers carry a confidence. When the router isn't sure, it escalates to your most capable tier instead of guessing — you set the bar.
More on routingSign up with your name and email.
One key per app or environment. Revoke any time.
Send your key as a Bearer token and act on the decision.
Create an account in a minute. Talk to us when you need higher limits.