Claude Code Writes It. Kane CLI Proves It Works.

What Claude Code Does Well, and the One Thing It Cannot

An AI coding agent reasons about code well. It reads your repo, follows your patterns, and produces a working diff. It can even guess whether the result looks right.

What it cannot do on its own is run the rendered app the way a user would and know, for certain, that the login set the session, the cart updated, and the page showed what it should. Looking right and being right are different things. The agent has no deterministic way to tell them apart.

That is not a prompt problem. It is a missing tool. The agent needs a real browser and a clear pass or fail.

Note: Give your agent a real browser and a deterministic pass or fail. Start free. Try Kane CLI

How Claude Code and Kane CLI Close the Loop Together

Give Claude Code one new tool and the loop closes. Claude Code writes the feature. It calls Kane CLI. Kane CLI opens a real Chrome browser, runs the flow described in plain English, and returns a structured pass or fail.

Claude Code reads that result. On a pass, it moves on. On a fail, it gets the reason, the failing step, and the evidence, then fixes the code and runs it again. The same agent that wrote the feature now verifies the feature, with no human in the middle.

Wiring Kane CLI into Claude Code in One Step

Install Kane CLI, then add its skill where Claude Code can find it.

npm install -g @testmuai/kane-cli
kane-cli login

Drop the Kane CLI skill into ~/.claude/skills/kane-cli/. Claude Code loads it automatically, and from then on it recognizes browser tasks and calls Kane CLI on its own.

For agent runs, Kane CLI speaks a format the agent can parse. Use agent mode and read the last line.

kane-cli run "go to https://yourapp.com, log in as '{{user}}', assert the dashboard shows 'Welcome'" --agent --headless

kane-cli run "..." --agent 2>/dev/null | tail -1 | jq .

Output is one JSON object per line. The final run_end line carries the status, a summary, any values you asked it to extract, the duration, and a link to the full run. Claude Code reads that line and decides what to do next.

Note: Want the full agent mode and skill setup? Read the Kane CLI docs. Read the docs

What Changes Once Your Agent Can Verify Its Own Work

The loop stops needing you in the middle. Claude Code can run a flow, read the result, diagnose a failure from the logs and screenshots, fix it, and rerun, all in one pass. It can even spin up parallel runs across sub-agents to check several flows at once.

You stop being the verification step. You review outcomes instead of every click. The agent ships code it has actually confirmed works, with a shareable link as proof.

Note: The Claude Code skill lives in the repo. Grab it and the examples to set it up. Open the repo

Claude Code already writes the code. With Kane CLI, it can prove the code works, in plain English, in a real browser, before it ever reaches you or a user.