The interview room
Watch them build the thing.
Not describe it on a whiteboard. Not type into a pad that cannot run. The call, the repository, a real terminal and a live view of the candidate’s agent, in one room on both machines at the same time — staged by the preset before anybody arrives.
A shared workspace, not a shared screen
Screen sharing turns one person into a spectator watching a compressed video of somebody else's font. The room mirrors the work itself, so both sides get a real, readable, local view of it.
Their editor, as they type
Content and caret stream continuously, not on save. You see where they are in the file, and following pulls you to whatever they open next — a toggle, because reading ahead while they think is a normal thing to want.
Their terminal, live
Their shells are mirrored on your machine as real terminals. Tests run and you watch them run. Each keeps a tail of scrollback, so joining late shows you how they got here rather than a blank panel.
Type in their terminal, if they say yes
Pairing on a stuck build is a real thing to want. It asks them first, and until they accept, nothing you type goes anywhere.
Screen share, as the backup
For everything that is not code — devtools, a simulator, a design file, a stubborn local setup. It gets a full pane rather than a thumbnail, because a shared screen you cannot read is the same as no shared screen.
Bring your own AI
Pretending an engineer does not use an agent is a test of who is willing to pretend. The interesting question stopped being whether they can write a function unaided and became how well they direct, read and correct a model — so the room is built to show you exactly that.
The candidate's own
They sign in to their own Claude — their subscription, on their machine. Nothing is billed to you and nothing is relayed through us. You see every prompt, tool call and result as it lands.
Or yours, scoped to the hour
If you would rather supply the agent, it relays through a session-scoped token that dies with the interview and never touches the candidate's machine.
Or none
Turn it off for the round. Some questions are better without one, and that should be your call rather than a limitation of the tool.
The same choice runs through the rest of the product: the console that moves candidates and builds loops runs on your own Claude when you have it installed, and on ours when you do not. It is a picker, not a detection you have to reverse-engineer.
Where the room comes from
Every round in every loop points at a preset, and the preset is what stages the room. Change it once and every interview at that stage changes with it.
Staged before anyone arrives
The preset attached to this round already knows what the room should look like, so the first ten minutes are the interview rather than setup.
Repositories
Cloned on both machines. Real files, branches and history — not a snippet in a text box.
Panes
The layout the round calls for: code, preview, tests, transcript, video, screen.
Terminals
Opened in the repository, running whatever the round says to run.
The brief
Pinned where the candidate can actually read it, instead of pasted into chat.
What you have afterwards
An interview nobody can reconstruct is an opinion. The room writes down what happened while it happens.
One transcript, one clock
Speech, agent activity and repository events on a single timeline, attributed to the person who did them.
Test runs
Pass and fail per test with the output underneath, so “it worked” is checkable.
Scorecards
Criteria live on the round, so two people who sat the same stage are comparable. Nobody sees another card before filing their own.
Integrity, without a spyware pitch
Remote interviews have a real problem and most answers to it are worse than the problem. This one asks the candidate's permission, in plain words, and tells them exactly what it will and will not look at.
If an interviewer turns the checks on, the candidate sees the full list before anything runs and can decline — and declining is recorded rather than punished silently. It never reads their screen, their keystrokes, their files or any other application’s contents, and it cannot stop anything from running. Findings reach the interviewer as context, not as a score.
The room is the easy part to judge.
Run one against a repository you know well. Ten a month, free, on the whole product.
Get started free