Skip to main content
Glama

agenteyes-mcp

A Node.js MCP server (stdio transport) exposing agent tools backed by headless Chromium. One tool call returns screenshot PNG(s), an accessibility snapshot, and page text — eyes for agents.

Requirements

  • Node.js 20+

  • On first install, Playwright downloads the Chromium headless shell (postinstall script)

Related MCP server: Browser Intelligence MCP

Install and run

npx agenteyes-mcp@0.1.0

Or from a clone:

git clone https://github.com/tech-lead-bunny/agenteyes-mcp.git
cd agenteyes-mcp
npm install
node index.js

The process speaks MCP over stdio; connect it from an MCP host (Claude Desktop, Cursor, etc.).

Tools

Tool

Description

screenshot_url

Open a public http/https URL; return PNG screenshot(s), accessibility tree, and visible text

get_page_text

Same page context without screenshots (title, a11y, text)

screenshot_url parameters

  • url (required) — public http(s) URL

  • full_page (optional, default false) — capture full scroll height

  • segment_height (optional, default 0) — when full_page is true and the page is taller than this value, capture stacked viewport PNGs of this height (top to bottom), then one combined accessibility + text block

  • width / height (optional) — viewport size (defaults 1280×720)

Security (SSRF)

URLs are checked before navigation and again on the final URL after redirects:

  • Only http: and https:

  • DNS resolution must not point at private/local addresses (localhost, loopback, RFC1918, link-local 169.254.0.0/16, fe80::/10, ::1, IPv4-mapped IPv6, etc.)

Claude Desktop

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "agenteyes": {
      "command": "npx",
      "args": ["-y", "agenteyes-mcp@0.1.0"]
    }
  }
}

Cursor

In Cursor Settings → MCP, add a stdio server:

{
  "mcpServers": {
    "agenteyes": {
      "command": "npx",
      "args": ["-y", "agenteyes-mcp@0.1.0"]
    }
  }
}

Or use a local path after npm install in this repo:

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

Example prompt

Use the screenshot_url tool on https://example.com with full_page: true. Summarize the page title and main message from the returned text, and describe what you see in the screenshot.

Limits (v0)

  • No hosted API — run the MCP server locally (or wherever your host spawns it)

  • No authentication or rate limiting in the server itself

  • Public URLs only — private networks and metadata endpoints are blocked

  • Chromium only — no Firefox/WebKit; no persistent login sessions between calls

  • One browser process, one page per tool call — pages are always closed after each call

  • Best-effort rendering — heavy SPAs, infinite scroll, or bot blocking may produce incomplete captures

  • Segmented full-page capture returns multiple PNGs (not a single stitched image)

MCP registry

  • mcp.so — primary listing target; submission to the registry is planned once 0.1.0 is on npm.

  • awesome-mcp — PR to add agenteyes-mcp planned the same week as the npm publish.

License

MIT — see LICENSE.

Available Tools

2 tools
get_page_textA

Open a URL in headless Chromium and return the page title, accessibility snapshot, and visible text (no screenshot).

Inputs: url (required).

screenshot_urlB

Open a URL in headless Chromium and return PNG screenshot(s), an accessibility snapshot, and visible page text.

Inputs: url (required), width, height, full_page, +1 more.

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 2 tools

    Disambiguation4/5

    The two tools are mostly distinct: screenshot_url includes screenshots, while get_page_text does not. However, both return accessibility snapshots and visible text, creating some overlap that could lead to arbitrary selection when only text/snapshot is needed.

    Naming Consistency3/5

    Both names use snake_case, but verb patterns are inconsistent: 'screenshot' acts as a verb/noun, while 'get_page_text' uses the verb 'get'. The mix is readable but not a predictable pattern.

    Tool Count3/5

    Two tools is thin for a web content extraction service. While each tool has a role, the surface feels minimal and could benefit from additional focused tools (e.g., raw HTML, links).

    Completeness3/5

    Core capabilities (screenshot and text/snapshot) are covered, but notable gaps exist: no way to retrieve raw HTML, page links, or interact with the page. Agents may hit dead ends for structured data extraction.

    Maintenance

    ActivitySlowing
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      D
      maintenance
      Enables agents to take screenshots and extract text from web pages using a persistent Chrome browser via CDP, with built-in SSRF protection.
      MIT
    • A
      license
      A
      quality
      C
      maintenance
      Enables AI agents to read any web page as clean Markdown, pull title/description/preview metadata, capture desktop, mobile, or full-page screenshots, and list every link or image on a URL. JavaScript-heavy pages are rendered in a real browser and blocked requests retried through a proxy, so agents can reach content they'd otherwise miss.
      5
      42 npm
      MIT