livetest — Claude Code Plugin

A Claude Code plugin that makes Claude test a feature in the actual running app — driving a real browser via the Playwright MCP server, inspecting live app state, verifying with evidence, and cleaning up after — instead of only reasoning about whether the code is correct.

App-agnostic: works for any web app, dashboard, or API. Not tied to any one project.

Installation

/plugin marketplace add MKToronto/livetest
/plugin install livetest@livetest

Then restart Claude Code (/exit followed by claude) to load the plugin.

For local testing without installing:

git clone https://github.com/MKToronto/livetest.git
claude --plugin-dir /path/to/livetest

Updating

/plugin marketplace update livetest

Requirements

This plugin drives a real browser through the Playwright MCP server, so that server must be configured in Claude Code. It provides the mcp__playwright__browser_* tools the command uses.

Add it once (user scope, available in every project):

claude mcp add playwright npx @playwright/mcp@latest

No separate browser-install step is needed — the server sets up its browser on first use. (If you ever need to install browsers manually: npx playwright install.)

Verify it registered:

claude mcp list

The app under test must also be running locally — Claude confirms it's listening before driving it and won't launch it for you unless you tell it to.

Usage

/livetest <what to test>

• /livetest the date-range filter — test a specific thing.
• /livetest (no argument) — test whatever was just worked on this session, inferred from recent edits + the conversation.

What it does

Claude follows a fixed playbook:

1. Find the running app — looks up the dev/start command and port, confirms it's listening, then navigates a real browser to it.
2. Reach the feature through the real UI — interacts with the real controls (or hits the API directly for backend work), exercising the same code path a user does.
3. Inspect live state — reads the app's real state (Svelte stores / Redux / exported getters / DOM / network), not just pixels.
4. Verify with evidence — asserts concrete values (counts, flags, response bodies) plus a screenshot and the console error log.
5. Clean up — undoes test state, cancels anything it started, removes screenshots and .playwright-mcp.
6. Report honestly — states what was verified with evidence, and explicitly what was not exercised.

The full playbook lives in commands/livetest.md.

Safety

livetest will not trigger irreversible or outward-facing actions against live systems — no real submissions, sent messages, payments, deletes, or writes to production — unless you've explicitly authorized that action for the session. Destructive or external steps are stubbed, fixtured, or reported short of execution.

License

MIT — see LICENSE.