Every applicant gets a real interview.

No, and it is built so that it cannot. Your knockout criteria are never sent to the interviewer, so it has nothing to judge against during the conversation. It reports what was said, with the sentence attached, and a person on your team decides. A disqualifier that was asked and never really answered comes back as unclear rather than cleared, because a missed borderline candidate costs you one interview and a wrong rejection costs someone a job.

On a coding assessment, yes, and we assume they will, because they will be using it on the job too. The assistant is in the editor we hand them, so every question they ask and every suggestion they take comes back to you with the code. Then they defend what they submitted out loud, with questions generated from their own lines. Someone who cannot explain the code they handed in runs out of answers in about ninety seconds.

The spoken half. A stand-in can write the code; the interview afterwards is live, unscripted and built from that specific submission. We do not run proctoring software, we do not watch their face and we do not score their eye movements. Documented false-positive rates for automated proctoring run 15 to 30% and fall hardest on neurodivergent candidates, which is a cost we are not willing to hand you.

It is built for the rules that already exist. Illinois AIVIA requires notice, an explanation and consent before AI evaluates an interview, NYC Local Law 144 requires notice, the EU AI Act treats hiring as high risk, and Quebec Law 25 governs the recording. Every candidate is told plainly what is happening before anything starts, consent is timestamped and auditable, and they can have their recording and transcript destroyed on request. Criteria that touch a protected characteristic are refused at the point you write them, with the protection named. No biometrics, no facial analysis, no affect scoring, ever.

Then you will be able to see it, which is the part that matters. Every number sits above the quote that produced it, and every quote is checked against the transcript before it reaches you. The report also carries a panel saying what it does not know: the disqualifier nobody answered, the load test that observes behaviour and cannot prove thread safety. On day one your knockout verdicts will still vary run to run, which is exactly why every large buyer starts on one role rather than a whole pipeline.

A screen is a link, so anything that can send a candidate a link already works, which is every ATS. Beyond that there is an MCP endpoint, so the assistant your team already uses can ask who is waiting on you, pull up what somebody actually said, and record a decision, without anyone opening the dashboard. Read only by default. There is no migration and nothing to rip out if you stop.

Your role and company, not our branding. A voice screen is whatever length you set, usually ten to fifteen minutes; a coding assessment is about seventy-five minutes to write and twelve to defend. They are told it is an AI interviewer, that a human decides, and that they can delete their recording. Unless you turn it off, they also get a short note on what went well and one thing to work on, never a score, because Larpy is an interview practice product first, which is the reason a candidate has any reason to think well of having been screened this way.

Per screen, on an annual term billed monthly, with an included volume and a metered rate above it. The rate falls as volume commits. What moves your number is how many you run, how many roles at once, and whether you want the coding assessment alongside the interview. Nothing is sold as unlimited, so what a contract costs to serve is known before it is signed. We quote on a call rather than publishing a ladder, because the right shape for four roles and for four hundred are not the same product, and almost everyone starts with one role for thirty days.

Paste a job posting and you have a screen you can edit and send in about a minute, which you can do on this page right now. A pilot is one role for thirty to forty-five days, and it exists so the auto-reject rate is proven on your funnel before it carries anything important.