Verified with Kane CLI

Your AI agent
says done.
Handoff makes
it prove it.

Write what "done" means in plain English. Kane CLI checks every line in a real browser, and Claude Code fixes what fails, until it's proven.

Sproutlocalhost:4173
Checks
Round 1FixRound 2Sweep
Kane readout · yearly-pricing
—waiting
0/6verified in Chrome
Claude Codeidle
The problem

"Done" is a claim. Someone still has to click through it and find what's broken.

Both bugs below are real, from the agent's first build of our demo app. Drag to compare.

Round 1: Kane typed ada@example and the page said You're on the list Round 2: the page shows Enter a valid email address Round 1 · accepted ada@exampleRound 2 · fixed
reject-bad-email

Kane typed ada@example. The page said "You're on the list."

yearly-pricing
$60$48

What Kane read off the page, and what the spec says.

One agent round

Claude Code fixed both and left the working code alone. Kane re-checked: 6/6.

How it works

Spec in. Proof out. No human in the middle.

01

Writethe spec

Checks in plain English, the way a designer writes acceptance criteria. Brand rules on top. No selectors, no test code.

handoff.mdPlain EnglishBrand rules
---
url: http://localhost:4173
---
## Checks
- yearly-pricing: Click "Yearly",
  verify the Grow plan shows "$48"
- reject-bad-email: Enter "ada@example",
  verify "Enter a valid email address"
02

Claude Codebuilds it

Your coding agent builds the app, or works in it as usual. Handoff never edits your code itself.

Claude CodeAny appAny stack
❯ Build the Sprout launch page from handoff.md
● Write(public/index.html)
● Write(public/app.js)
Done. The waitlist, pricing toggle and FAQ are in.

# "done"… but is it?
03

Kane CLIproves it

Every check runs in a real Chrome browser through Kane CLI's agent mode. Pass or fail, with exactly what it saw and a link to the run.

kane-cli run --agentReal ChromeNDJSON
◉ scroll: Scrolling to the Simple pricing section
◉ click: Clicking the Yearly button
✖ FAIL The Grow plan shows $60 per year,
  not the required $48.
  Kane read "$60" · expected "$48"
04

Handoffcloses the loop

Failures become a precise brief for the agent. It fixes, Kane re-checks what failed, then sweeps every check again. Until it's green.

handoff loopStop hookRegression sweep
── Agent · fixing 2 failing checks
◆ read public/app.js
◆ edit public/app.js
── Round 2 · Kane re-checks what failed
✔ PASS selected yearly pricing
✔ Verified. 6/6 after 1 agent fix round.
Kane says why

Kane says why it failed. Handoff listens.

Kane CLI classifies every failure. So the agent only ever gets sent after real bugs.

The app is wrong

Brief the agent with the evidence

The values Kane read, its verdict, the brand rules, and a link to the run.

objective_contradiction
read "$60" · expected "$48"Fix
Kane's own click slipped

Re-run before touching any code

A flaky click isn't a bug in your app, so nobody "fixes" working code.

automation_bug · agent_misstep
Re-running checkRetry
The browser itself failed

Report it. Never blame the code.

Screenshot timeouts and crashes are environment problems, not app bugs.

infra_error.screenshot_failed
Environment, not your codeSafe
Proof, not logs

Send a proof page, not "trust me."

The proof pageThe designer's words on top, Kane's evidence underneath, the loop round by round.
Kane's verdictsRound 1 on a real Windows machine: two failures, with the exact reason.
The agent's fixClaude Code reads and edits app.js, then Kane re-checks only what failed.
Round 2 in ChromeThe same email, now rejected: "Enter a valid email address."
Two ways in

Stay in Claude Code, or let it run on its own.

The Stop hook

handoff init adds a Claude Code Stop hook. Claude can't call it done until Kane agrees. A failure becomes its next instruction.

❯ Add a FAQ question "Is there a student discount?"
● Update(public/index.html) · Added 1 line
✻ Handoff: verifying the spec in a real browser…
Stop says: Handoff ✔ 6/6 checks verified by Kane CLI

The headless loop

handoff loop runs on its own: verify, brief Claude Code, re-check, sweep. Locally or in CI.

── Round 1 · Kane CLI verifies the build
✖ FAIL The Grow plan shows $60 per year, not $48.
◆ edit public/app.js
✔ Verified. 6/6 after 1 agent fix round.
Use it on your project

Write what "done" means once. Kane checks it every time.

Think of it as a QA teammate for your coding agent. You stop being the one who clicks through and catches its mistakes.

01 · Once per computer

Install the tools

Node 18+, Google Chrome and Claude Code, plus Kane CLI with a free TestMu AI account. Then Handoff itself.

npm i -g @testmuai/kane-cli
kane-cli login
npm i -g github:esteexbt/handoff
02 · Once per project

Set it up in your app

Run this in your project folder. It creates handoff.md and connects Handoff to Claude Code.

cd my-app
handoff init --url http://localhost:3000
03 · Write what "done" means

Plain English, one line per check

How to start the site, then what someone should see when they use it. The way you'd brief a developer.

## Checks
- signup: Submit the form with "ada@example",
  verify it shows "Enter a valid email"
04 · Then just build

Claude can't say done until Kane agrees

Work with Claude Code as usual. Each time it finishes, Kane tests your checks in Chrome. Anything broken goes straight back to Claude to fix.

✻ Handoff: verifying in a real browser…
Stop says: Handoff ✔ 6/6 checks verified
Run a quick check, any time
handoff verify
Let Handoff fix and re-check on its own
handoff loop --open
Get started

Two minutes to your first proof.
Open source. Zero dependencies.

$ npm i -g github:esteexbt/handoff
01handoff initCreates handoff.md and the Claude Code Stop hook.
02handoff verifyEvery check, in a real browser, with Kane CLI.
03handoff loop --openClaude fixes what Kane catches. The proof page opens.
View on GitHub ↗