Spandan Mishra

Codelit (https://codelit.duckdns.org) is an AI-first coding assessment platform built for a world where every candidate already has an agent. The problem: traditional coding tests grade the final answer — and a photo of a problem plus an AI model produces a perfect final answer. Green ticks now prove nothing about whether a person understands code, can direct an AI agent, or could maintain what was submitted. Codelit grades what a screenshot cannot produce: how you drive an agent under constraints, and whether you can defend the result under time pressure. How it works: 1) Bring your own agent — pair your computer once with a one-time code via a small local connector (codelit-connect). Your provider login never leaves your machine; Codelit only receives a revocable device credential, your prompts, tool activity, and solution revisions. Codex and Muse Code are both supported. 2) Solve staged scenarios — requirements arrive in three stages with a prompt budget. The platform brief is never handed to the agent, so the candidate must understand and explain the task themselves. Every prompt and file change is captured automatically as a trace. 3) Finish — the solution is frozen under a content hash and evaluated server-side against visible checks plus hidden acceptance tests (only pass counts are revealed). 4) Defend the trace — a timed dialog (90 seconds per answer) asks three personalized questions: a quoted line from your exact code, a design trade-off, and the prompt decisions behind the run. 5) Bank your passport — only defended ranked runs become a credibility record: score, tier, prompt count, and defense performance. Ranked unlocks after five distinct practice passes; scenarios are generative (seeded variants), so memorized or shared solutions do not transfer. Candidate code runs locally in disposable sandboxed namespaces (CPU/memory/process/network limits) — never via a model. In short: Codelit doesn't ban AI. It makes cheating pointless.