MCP server

AI Agent Benchmarking for Websites

Help Claude browse your website better. Serge MCP instruments every navigation step — clicks, searches, form fills — with timing, screenshots, and detailed accessibility analysis so you can optimize your site for AI-assisted browsing.

What Serge MCP does

When Claude browses a website through Serge MCP, every action is captured: how long pages take to load, whether buttons and form fields are discoverable, where navigation breaks down, and why. After a session, Serge MCP generates a detailed report with a step-by-step timeline, screenshots, and actionable findings.

Serge MCP runs as an MCP server inside Claude Desktop. No API keys, no separate accounts, no billing — it works on your existing Claude subscription.

Quick start

Add Serge MCP to your Claude Desktop configuration file.

macOS ~/Library/Application Support/Claude/claude_desktop_config.json
Windows %APPDATA%\Claude\claude_desktop_config.json
{
  "mcpServers": {
    "serge": {
      "command": "npx",
      "args": ["-y", "@serge-ai/mcp-server"]
    }
  }
}

Restart Claude Desktop. On first use, Serge MCP automatically installs the browser it needs — no manual setup required.

Open Claude Desktop and type:

Use Serge to find the cheapest wireless headphones on digitec.ch

Claude will start a benchmarking session, open a browser window, and navigate the site while Serge MCP captures everything. When the task is done, the report opens automatically in your browser.

You can also view past reports from the terminal:

npx @serge-ai/mcp-server report

How it works

Serge MCP provides Claude with a set of browsing tools: navigate, click, type, scroll, read page, and take screenshot. When Claude uses these tools instead of its built-in browser, Serge MCP captures telemetry on every action.

Each action records: the element targeted, whether it was found, how long the interaction took, the resulting page state, network requests fired, and a screenshot of the page after the action.

At the end of a session, all captured data is compiled into a single self-contained HTML report you can open in any browser, share with your team, or archive.

Tools

Serge MCP exposes eight tools to Claude Desktop.

serge_start_session

Begins a benchmarking session. Requires a target domain and a task description. Must be called before any other browsing tool.

serge_navigate

Opens a URL in the browser. Returns the page title and an accessibility tree so Claude can understand the page structure.

serge_read_page

Returns the current page's accessibility tree. Useful when Claude needs to re-examine the page after scrolling or waiting for content to load.

serge_click

Clicks an element identified by its accessibility role and name. For example: role "button", name "Add to cart". Returns whether the element was found, what tag it was, and whether the click triggered a navigation.

serge_type

Types text into a form field identified by its accessibility role and name. Can optionally press Enter after typing, which is useful for search fields.

serge_scroll

Scrolls the page up or down. Useful for finding content below the fold or triggering lazy-loaded elements.

serge_screenshot

Takes a screenshot of the current page and returns it to Claude as an image. Use when Claude needs visual context beyond the accessibility tree.

serge_end_session

Ends the session and generates the report. Claude provides an outcome (success, failure, or partial) and notes on what happened.

Example prompts

Copy these directly into Claude Desktop. Note: in conversation, just say "Use Serge" — Claude knows which tools to use.

E-commerce product search

Use Serge to search for "running shoes" on zalando.ch and add the first result to the cart.

Form completion

Use Serge to find the contact form on example.com and fill it out with test data.

Navigation depth

Use Serge to navigate from the homepage of digitec.ch to a specific product category (laptops) and find a product under 1000 CHF.

Checkout flow

Use Serge to add any item to the cart on brack.ch and proceed as far as possible through checkout without entering payment details.

Content discovery

Use Serge to find the return policy on galaxus.ch starting from the homepage.

Reading the report

The report is a single HTML file that opens automatically in your browser when a session ends. It contains these sections.

SummaryTask description, target domain, outcome, total steps, and total time.
Step-by-step timelineA vertical timeline showing every action Claude took. Each step includes the action type, parameters, duration, and a screenshot thumbnail you can click to expand. Failed steps are highlighted.
FindingsAutomatically generated insights from the session data. These include: elements that took multiple attempts to find, actions that failed entirely, pages that loaded slowly, bot detection events, missing ARIA labels, and inaccessible content like iframes or shadow DOM.
Network summaryTotal requests, page weight, blocked requests, and the slowest-loading resources.

Common findings and what they mean

Missing ARIA labels

Claude couldn't find an element by its accessible name. The element may exist visually but lacks the semantic markup that makes it discoverable to AI agents. Fix by adding aria-label attributes to interactive elements.

Bot detection blocked

The site's CDN (Cloudflare, Akamai, etc.) served a challenge page instead of the actual content. This may prevent AI agents from accessing your site. Consider adding agent-friendly access paths or whitelisting known AI user agents.

Element not found

Claude tried to interact with something that doesn't exist in the accessibility tree. Common causes: the element is rendered via JavaScript after page load, it's inside an iframe, or it uses a non-standard component that doesn't expose accessibility information.

Slow page load

A page took more than 3 seconds to reach a usable state. AI agents are sensitive to load times because each navigation step adds to the total task duration.

Iframe content inaccessible

Forms or content embedded in iframes are often invisible to the accessibility tree. If critical user flows (like checkout or payment) rely on iframes, AI agents may not be able to complete that step.

Duplicate accessible names

Multiple elements share the same role and name (e.g. several "Add to cart" buttons on a product listing page). Agents cannot distinguish between them and may interact with the wrong element. Fix by adding unique aria-label attributes that include product context.

Track progress over time

Want to see how your site's agent readiness improves? Upload your session reports to the Serge dashboard to track changes, compare runs, and share findings with your team.

Save to Serge Dashboard

Viewing past sessions

List all previous benchmark sessions:

npx @serge-ai/mcp-server history

Open a specific report:

npx @serge-ai/mcp-server report <session-id>

Session data and reports are stored locally at ~/.serge/.

Configuration

Serge MCP works out of the box with no configuration. The browser launches in headed mode (visible window) so you can watch Claude navigate in real time.

All session data is stored locally on your machine at ~/.serge/sessions/. Reports are saved at ~/.serge/reports/. Nothing is sent to external servers.

Requirements

Node.js 18 or later. Claude Desktop with an active Claude Pro or Max subscription. macOS or Windows.

Troubleshooting

Tools don't appear in Claude Desktop

Make sure you fully quit Claude Desktop (Cmd+Q on Mac, not just close the window) and reopen it after editing the config file. Check that the JSON syntax is valid. Look for errors in ~/Library/Logs/Claude/mcp*.log on macOS.

Browser doesn't open

On first run, Serge MCP installs Chromium automatically. This can take up to a minute. If it fails, run npx playwright install chromium manually in your terminal.

Claude doesn't use Serge's tools

Be explicit in your prompt: mention "Use Serge" or reference a specific tool like serge_start_session. Claude may default to its built-in browser if the prompt is ambiguous.

Bot detection blocks the session

Some sites aggressively block headless browsers. Try running with a headed browser (the default) and avoid rapid sequential requests. Bot detection is itself a finding — it may mean AI agents are not able to use the site.

Accessibility tree is too large

On complex pages, the tree may be truncated to fit within Claude's context window. Serge MCP handles this automatically by trimming the tree and noting the truncation. Use serge_scroll to explore different parts of the page.

Privacy

Serge MCP runs entirely on your machine. Session data, screenshots, and reports are stored locally at ~/.serge/ and are never transmitted to external servers. The only network traffic is between the local browser and the target website you choose to benchmark.

Serge MCP does not collect analytics, telemetry, or usage data. There is no account registration or authentication required.

Full privacy policy: serge.ai/privacy

Support

GitHubhttps://github.com/SuperstellarLLC/serge-mcp-server
Issueshttps://github.com/SuperstellarLLC/serge-mcp-server/issues
Emailsupport@serge.ai
Websitehttps://serge.ai

MIT — Superstellar LLC