Interviews that watch the work, not the webcam.
Workbench gives every candidate a zero-setup workspace in the browser — a private sandboxed machine with a real terminal, Claude Code preinstalled, and a GPT chat panel. AI use is encouraged and evaluated. Every prompt, diff, and command is captured by construction, and the whole session replays for a human to judge.
From your repo to a decision — in five steps.
Connect your repo
Point Workbench at your GitHub repo — or upload a zip, or start from a role preset. Challenges are generated from code that only exists at your company, so they look like the job and can't be Googled.
Generate the challenge
The AI drafts a task scoped to be completable in your time budget, with visible checks the candidate can run themselves. No hidden tests, no trick questions — you review and adjust before anything goes out.
Invite the candidate
They click a link and land in a ready workspace on a private sandboxed machine — nothing to install, no environment to fight. AI is on the desk from the first minute, and using it well is part of the evaluation.
Watch the replay
Every prompt, diff, terminal command, and timing signal is captured by construction — there's no route around it. One fused timeline replays exactly how they worked, with evidence chips that jump to the moment behind every score.
Decide
Async submissions auto-score the instant they land, and a human can supersede any number before it counts. One click turns a submission into a live follow-up, with a prep kit brewed from the candidate's actual work.
Built so the evidence is the product.
Zero-setup sandbox with Claude Code
One link opens a private machine in the browser: real terminal, real filesystem, Claude Code preinstalled, and a GPT chat panel. The machine exists only for the session and is torn down after.
Cheat-resistant by construction
Challenges are generated from your own codebase and scoped to the time budget. There's no Stack Overflow answer for code that only exists at your company — and every check is visible, so candidates know exactly what done means.
Every prompt and diff, captured
The workspace's only network route is the audited proxy, so capture isn't monitoring bolted on — it's how the machine works. Prompts, diffs, commands, timing. No webcam, no screen recording.
One fused replay timeline
Prompts, diffs, terminal output, and check runs land on a single timeline. Scrub to any moment and see what the candidate asked, what the AI produced, and what they did with it.
Instant scores a human supersedes
Async take-homes auto-score against the visible checks the moment they're submitted. Reviewers get evidence-linked scorecards — and the human call always outranks the machine's.
From submission to live interview
Any submission can turn into a live follow-up, with a prep kit brewed from the work itself — the diff they wrote, the prompts they used, the check they never got green. Your interview, amplified.
The best engineers direct AI now. Interview for that.
Every serious engineering team works with AI on the keyboard — except in interviews, where we confiscate the tools and test people on a job that no longer exists. Workbench flips it: give candidates the real tools, watch how they direct them, and judge the judgment. There's no AI interviewer persona here — async challenges are AI-guided work samples a real human reviews, and live interviews stay yours. The AI does the work. You make the calls.
The work is the interview.
Watch it happen.
Book a 20-minute demo and see a challenge generated from your own repo — then scrub the replay. Included in existing Chosen plans.