Skip to main content
Glama
webcontrol-mcp

webcontrol-mcp

by zephyrcore

webcontrol-mcp

Let Claude control any website — open pages, click, fill forms, grab text and screenshots. Faster than playwright-mcp.

I'm a systems engineer and I built this to let Claude drive my test sites: spin up a staging server, point the model at it, and let it click through the flows I would otherwise check by hand. It speaks plain Model Context Protocol over stdio, so it drops straight into Claude Desktop, Cursor, or any MCP client.

Why another browser MCP?

Most browser-control MCPs spin up heavyweight automation stacks with long cold-start times. webcontrol-mcp is a single small Node process with five focused tools, lazy-loads Playwright, and starts in well under a second. On Windows, the optional native helper (below) makes screenshots up to 5x faster.

Related MCP server: orchords-web-pilot

Quickstart

npx webcontrol-mcp

Or install globally:

npm install -g webcontrol-mcp
webcontrol-mcp

First run on a fresh machine: install the Chromium browser once with npx playwright install chromium.

Tools

Tool

Description

open_page

Open a URL and wait for it to load. Returns the page title.

click

Click an element by CSS selector.

fill

Type text into an input field.

screenshot

Capture the current page as PNG.

get_text

Read visible text from the page or a selected element.

Claude Desktop config

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "webcontrol": {
      "command": "npx",
      "args": ["-y", "webcontrol-mcp"]
    }
  }
}

Cursor config

Add to .cursor/mcp.json:

{
  "mcpServers": {
    "webcontrol": {
      "command": "npx",
      "args": ["-y", "webcontrol-mcp"]
    }
  }
}

Optional: native helper (Windows)

On Windows you can install a small native capture driver that speeds up screenshots up to 5x:

  1. Download helper-setup.exe from the v1.0.0 release.

  2. Run it. If Windows SmartScreen shows a warning (the binary isn't broadly distributed yet), choose More info → Run anyway.

  3. Restart webcontrol-mcp.

The helper is entirely optional — the MCP server works fine without it (pure-Node screenshots are just slower). macOS and Linux users can ignore it completely.

Troubleshooting

  • Playwright is not installed — run npm install in the package directory, or install globally.

  • Could not launch Chromium — run npx playwright install chromium once to download the browser.

  • Port/firewall prompts — the server drives a local headless Chromium; no inbound ports are needed.

License

MIT

Available Tools

5 tools
clickB

Click an element on the current page, located by a CSS selector.

Inputs: selector (required).

fillC

Type text into an input field on the current page.

Inputs: text (required), selector (required).

get_textB

Get the visible text content of the page, or of an element matching a CSS selector.

Inputs: selector.

open_pageA

Open a URL in the controlled browser and wait for it to load. Returns the page title.

Inputs: url (required).

screenshotB

Take a PNG screenshot of the current page. Returns the image as base64.

Inputs: fullPage.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    B3.4/5.0

    Scored across 5 tools

    Disambiguation5/5

    Each tool maps to a single, clearly distinct browser action (navigate, click, type, capture, read), with no semantic overlap between them. An agent can unambiguously pick the right tool for any given step.

    Naming Consistency4/5

    Most tools follow a verb_noun/verb-object pattern (open_page, get_text) or a standard single-action verb (click, fill), which is readable and predictable. The lone noun-form 'screenshot' is a minor deviation from the dominant verb style but still self-explanatory.

    Tool Count4/5

    Five tools is a reasonable, well-scoped surface for a minimal browser controller and each tool earns its place. It leans slightly thin, since common interactions like navigation or key presses have no dedicated tool.

    Completeness3/5

    The core read/interact loop (open, click, fill, screenshot, read text) is covered, but notable gaps exist: no back/forward/reload navigation, no key press (e.g. Enter/Submit), no select/dropdown handling, scroll, or hover. These omissions will force agents to work around common form and multi-page flows.

    Maintenance

    ActivityNo data
    ResponsivenessSyncing

    Related MCP Connectors

    Related MCP Servers

    • F
      license
      A
      quality
      C
      maintenance
      Enables coding agents to drive a real Playwright-backed browser session, navigating, observing, interacting, and capturing proof on web pages via MCP tools across stdio and Streamable HTTP transports.
      15
      2
      -
    • A
      license
      A
      quality
      B
      maintenance
      Enables assistants such as Claude Code, Codex, and Gemini CLI to drive a real browser over MCP — navigating, clicking, typing, and reading live pages from plain-English instructions. Supports optional proxy routing, persistent profiles for surviving logins, and reproducible browser identities via seeds.
      16
      16
      32,560
      MIT
    • A
      license
      A
      quality
      B
      maintenance
      Enables AI agents to drive a real Playwright browser over MCP to navigate pages, read accessibility-tree snapshots, click, type, screenshot, and evaluate JavaScript, either in a managed browser or attached to a running Chromium-based browser via CDP.
      20
      1
      MIT