The Kane CLI Hackathon Winners: Three Projects That Made Verification Real

🥇 demo.studio by Sam Felix

Lane: Browser agents in the wild

Every builder knows the last-minute scramble. The feature is done, but someone still has to click through it on camera so everyone else can see it. Sam built demo.studio to kill that chore.

demo.studio takes a live website and a short brief, the URL, the goal, the audience, the on-screen actions. It returns a narrated product demo video. No screen recorder. No retakes.

Why the Kane CLI integration stood out

Sam did not use Kane CLI once. He used it three ways.

First, as a build partner. The team built demo.studio in Cursor with recursive feedback from Kane CLI, so every iteration got checked in a real browser before moving on.

Second, as the performer. Kane is the agent that walks through the third-party product and films the walkthrough. The browser agent is not a gimmick here. It is the product.

Third, as the tester. Kane verifies demo.studio itself end to end: landing, wizard, gallery, health. The tool that films demos is itself demo-proof.

That layered use is what the judges rewarded. Kane CLI was not bolted on. It was the build loop, the runtime, and the safety net at once.

Repo: github.com/SamFelix03/demo.studio · Demo: Watch the video

🥈 GuardianKane by Abhinav

Lane: Verification baked into your workflow

Claude Code says "done." Is it? GuardianKane makes "done" a state the agent has to earn, not self-report.

GuardianKane is a Claude Code extension built entirely with Claude Code. It turns a PRD into a reviewed, phased task tracker. Every extracted use case and generated test has to pass kane-cli context review --approve before it counts as real.

Then comes the gate. When Claude marks a task as claimed done, a Stop-hook fires. It drives kane-cli to open a real local browser, replay a generated test, and run a second ad-hoc defect sweep against that task's PRD section. Three attempts, then it escalates to a human instead of looping forever.

Around that core sit a real-time scope guard that flags files touched outside a task's declared PRD scope, a secret-scan gate, file-level task locking, decision memory across sessions, and a live dashboard with PRD drift panels, per-criteria review cards, and a trace timeline per task. PRD changes sync through kane-cli maintain reconcile.

Why the Kane CLI integration stood out

Kane CLI is what turns every one of those checks from a claim into evidence. A companion demo repo shows the full closed loop: Kane catches a real checkout bug, the agent fixes it, Kane re-verifies. And four earlier experiments proved the gate catches what an unassisted agent ships silently: defects, doubled chart lines, layout deviations, a missing feature.

This is the pattern we built Kane CLI for. The agent builds it. Kane CLI proves it works.

Repo: github.com/18Abhinav07/adventures-with-kane · Demo: Watch the video

🥉 KANE vs. ABLE by Christine Lall

Lane: Apps that verify themselves

The most creative entry of the hackathon turned verification into a game. Literally.

KANE vs. ABLE is an adversarial AI escape room. ABLE, the Artificial Builder of Labyrinthine Escapes, creates a structured browser dungeon and claims it is solvable. Kane independently enters through Chrome, explores the room, gathers clues and items, and tries to reach the exit. Nobody hands Kane the solution.

The verified run tells the whole story. Kane failed the first dungeon. The failure evidence exposed a circular dependency in the room. ABLE repaired the dungeon through a coding agent, without touching Kane's objective. Kane automatically re-entered, independently derived the escape code 7294, entered it, and verified the win: YOU ESCAPED.

Why the Kane CLI integration stood out

This is a closed verification loop in its purest form. Builder claims. Verifier tests. Failure produces evidence. Evidence drives the fix. Verifier confirms the fix. The UI records everything: the failed and successful attempts, before and after dependency graphs, repair history, the discovered code, and ABLE's confidence throughout.

Kane was not scripted through the dungeon. It solved it. That independence is the point of plain English browser verification, and Christine made it playable.

Repo: github.com/christinelall/kane-vs-able · Demo: Watch the video

The Pattern Behind All Three

Three very different projects. One shared insight.

None of them treated Kane CLI as a test runner sitting at the end of a pipeline. Sam made it the runtime of his product. Abhinav made it the gate his coding agent cannot talk past. Christine made it an independent player that either escapes the room or does not.

In every case, the output is the same thing: a pass or fail backed by evidence from a real browser. Not a claim. Proof.

That is the quality gap AI agents leave behind, and 69 teams just showed us how many ways there are to close it.

Build Your Own Verification Loop

Everything the winners used is free to start. Install, describe what you want to test in plain English, and get a pass or fail with shareable proof in under 2 minutes.

npm install -g @testmuai/kane-cli