Quick start

Add the Hover MCP to the coding agent you already run, then ask it to test your app. No changes to your app, bundler config, or dependencies.

Prerequisites

A coding agent that speaks MCP — Claude Code, Cursor, or similar, installed and logged in. Hover rides whatever model it runs; no new API keys.
A dev server you already run — any framework Hover can reach over CDP (Vite, Next.js, Nuxt, Astro, webpack).

Don't have a coding agent yet?

Claude Code — npm install -g @anthropic-ai/claude-code, then claude login. Uses the Claude Pro / Max subscription you might already pay for.
Cursor — any MCP-capable agent works; add @hover-dev/mcp as an MCP server in its config.

Add the MCP

Install the MCP globally, then register the bin (Claude Code shown — any MCP-capable agent works):

npm i -g @hover-dev/mcp
claude mcp add hover -- hover-mcp

That's the whole install. Hover runs as an MCP server under your agent — no plugin in your project, no bundler-config edit, no new dependency, no keys of Hover's own.

Start your dev server

Use whatever command you already use:

pnpm dev          # or `npm run dev`, `yarn dev`, `bun dev`

Hover drives your existing dev server over CDP — it doesn't replace, wrap, or restart it.

First-run debug Chrome

On your first run, Hover prompts you to launch an isolated debug Chrome on port 9222 (separate from your everyday browser — a clean profile under <tmpdir>/hover-chrome). Subsequent runs reuse it. Hover never touches your normal Chrome session.

Ask your agent to test the app

In your agent, run the Hover tool:

/mcp__hover__test_app           # explore the app and crystallize a suite
/mcp__hover__test_app login     # …or scope it to one flow

The agent explores your app through Hover's grounded browser tools — click_control / fill_control and friends, which target controls by role+name so the selector that drives an action is the exact one saved. When a flow is verified, Hover crystallizes it.

The artifact

Specs land in __vibe_tests__/ as a <slug>.spec.ts file — plain @playwright/test code, with no LLM writing the code (crystallization is deterministic). It checks into git and runs in CI with no Hover dependency.

That's the loop: ask once → agent explores → crystallize → CI replays forever, with record == replay fidelity.

Run the specs in CI

The saved specs are plain Playwright. Run them anywhere with no AI, no tokens, no key:

npx playwright test __vibe_tests__

Point them at any environment with BASE_URL; the same specs run against local, staging, or a PR preview. Hover can generate a GitHub Actions workflow that runs them on every PR.

What's next

Your first session — A guided walkthrough on a real flow.
Pick an agent — Which coding agents work and how Hover plugs in.
Want a visual surface? The optional VS Code cockpit adds a Business Map graph + Dashboard. It reviews — it doesn't drive the agent.