visual-inspector-mcp

English | فارسی

An MCP server that lets Claude Code — or any MCP-compatible client — actually see and interact with the web pages it's working on, instead of only reading the source. It runs a persistent headless Chromium browser (via Playwright) and returns screenshots as inline images the model can view directly in the tool result.

"This icon looks wrong" → the agent screenshots the icon and looks at it, instead of guessing from the CSS.

Why

Coding agents are excellent at reading and writing code, but blind to what that code actually renders as. A misaligned icon, a color that doesn't match the design, a responsive layout that breaks at a certain width, a modal that overlaps its own close button — none of these are visible from source alone. This server closes that gap with a set of small, composable tools built around a single persistent browser session — including full form interaction, so the agent can fill inputs, type keystrokes, and submit forms before looking at the result.

Related MCP server: Browser MCP

Tools

Full parameter reference is in the tool descriptions themselves (visible to any MCP client) and in index.js.

Requirements

• Node.js 18 or later

• ~200 MB free disk space (Playwright downloads a Chromium build on install)

Installation

git clone https://github.com/MatinMHF/visual-inspector-mcp.git
cd visual-inspector-mcp
npm install        # installs dependencies and downloads Chromium via postinstall

Configuration

Claude Code

claude mcp add --scope user visual-inspector -- node /absolute/path/to/visual-inspector-mcp/index.js

--scope user registers it for every project. Use --scope project instead to scope it to the current repo, or omit --scope for the current session only. Restart Claude Code (or start a new session) after adding it — a running session won't pick up a newly registered server.

On Windows PowerShell, quote the -- separator:

claude mcp add --scope user visual-inspector "--" node "C:/path/to/visual-inspector-mcp/index.js"

Verify it's connected:

claude mcp list

Claude Desktop (or any client using claude_desktop_config.json)

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "visual-inspector": {
      "command": "node",
      "args": ["/absolute/path/to/visual-inspector-mcp/index.js"]
    }
  }
}

Other MCP clients

This is a standard stdio MCP server — node index.js speaks MCP over stdin/stdout. Any client that supports stdio-transport MCP servers can run it the same way.

Usage

Once configured, just ask naturally:

"Navigate to localhost:3000/settings and screenshot the save icon in the toolbar — does it look right?"

"Resize to 375x667 and check what the mobile nav looks like after I click the hamburger menu."

"Fill in the signup form with test data and submit it — any errors?"

"Any console errors on the checkout page?"

Typical tool sequences:

// Visual inspection
navigate({ url: "http://localhost:3000/settings" })
screenshot({ selector: "#save-icon" })   // isolate just the element
console_logs({ level: "error" })         // check for related JS errors

// Form interaction
navigate({ url: "http://localhost:3000/signup" })
fill({ fields: [{ selector: "#name", value: "Ada" }, { selector: "#email", value: "ada@example.com" }] })
press({ key: "Enter", selector: "[type=submit]" })
screenshot({})                           // see the result

// Responsive check
resize({ width: 375, height: 667 })
screenshot({ fullPage: true })

Design notes

• One persistent page per server process. navigate doesn't spin up a new browser each call — it reuses the same page/context so state (login, scroll position, SPA client state) survives across click → fill → screenshot sequences.

• fill vs type. fill is almost always the right choice — it clears the field, sets the value atomically, and fires the input/change events React's controlled components need. Use type only when the input's handler specifically reacts to individual keystrokes (e.g. an autocomplete that fires on keydown).

• Cheapest-capture-first. screenshot's tool description actively steers the calling model toward selector (one element) over the default viewport, and viewport over fullPage (most expensive in image tokens).

• Lossless by default. Screenshots are PNG unless you explicitly opt into format: "jpeg". On flat-color UI (icons, buttons), JPEG is often no smaller — so PNG stays the default; on photo-heavy full-page captures, JPEG at quality 80 measured ~75–80% smaller in testing.

• Bounded log output. console_logs defaults to the last 30 entries with no timestamps — pass limit/withTimestamps for more.

Security note

This server gives the connected AI client the ability to navigate a real (headless) browser to any URL it's given — including localhost and other addresses on your local network. It's designed to run locally over stdio for trusted development use, the same way you'd trust any other local dev tooling. Don't expose it over a network transport or hand it to an untrusted client.

Development

npm test        # runs smoke-test.mjs: spins up the server as a subprocess and
                # exercises every tool + parameter over the real MCP protocol

smoke-test.mjs is also run in CI on every push/PR (see .github/workflows/test.yml).

Changelog

v1.1.0

• Added fill tool — set input/textarea/select values (supports batch fields[])

• Added type tool — per-keystroke typing for inputs needing real keydown events

• Added press tool — send keyboard keys (Enter, Tab, Escape, Control+A, …)

• Extended smoke test with full form interaction coverage

v1.0.0

• Initial release: navigate, screenshot, click, resize, console_logs

License

MIT