MCP SERVER + BROWSER EXTENSION

Give Claude eyes on the tabs you already have open.

Scrapeline reads, extracts, and crawls the pages in your real, logged-in browser — through small, token-cheap MCP tools instead of screenshots. Everything runs on your machine. Nothing is uploaded anywhere.

runs in Chrome, Edge & Brave local-first, no account open protocol — MCP
shop.example.com/catalog
page_extract →
25 rows × 4 cols
saved catalog.csv (2.1KB)
HOW IT WORKS

One bridge, your machine, three hops.

Claude never touches your browser directly. Every call goes through a local WebSocket bridge that only your own extension can speak to — a web page can't reach it, and neither can anything outside your computer.

MCP client
Claude
stdio →
local process
MCP server
ws://127.0.0.1 →
your browser
Extension → Tab
CAPABILITIES

Escalate cheaply, extract precisely.

Claude starts with the cheapest tool that answers the question — a tab list costs a few tokens, a full crawl only runs when the job calls for it.

Look

Outline & read

page_outline · page_text · page_query

Structure, readable text, or a targeted selector read — each with its own token budget.

Extract

One page

page_extract

Auto-detects the main table or repeating list, or use your own selectors. Big tables save straight to a file.

Crawl

A whole listing

crawl_start

Follows Next links, a URL template, or infinite scroll. Survives the browser restarting mid-crawl.

Bulk

A list of URLs

extract_urls

One row per page from an explicit list — product pages, profiles, listings — no pagination needed.

Discover

Contacts & images

find_contacts · find_images

Emails, phones, social links, and images — on one page or across many, following internal links if asked.

Map

Sitemap & Shopify

sitemap_explore · shopify_catalog

Public, unauthenticated, instant — no tab needed. Reads a site's own sitemap.xml or product catalogue directly.

Repeat

Save a setup

recipe_save · recipe_run

Name any call above and replay it later, or hand the recipe file to a colleague running the same server.

Act

Scroll & open

page_scroll · open_url

Trigger lazy-loading, or open a URL in a background tab without stealing your focus.

Orient

What's open

list_tabs · page_info · browser_status

The cheapest possible look: which tabs are open, and whether the bridge is even connected.

MCP — AN OPEN PROTOCOL

Not just Claude.

Scrapeline speaks the Model Context Protocol, so anything that speaks MCP back can use it — the same local server, no extra work on its side.

CCClaude Code CLI · already configured

The repo ships a project-scoped .mcp.json — just run Claude Code from the project folder and approve it. From anywhere else:

claude mcp add scrapeline -- node /path/to/simple-web-crawler-scraper-extension/mcp-server/dist/server.mjs
CDClaude Desktop config file · one-time

Add a block to claude_desktop_config.json:

{ "mcpServers": { "scrapeline": { "command": "node", "args": ["/path/to/simple-web-crawler-scraper-extension/mcp-server/dist/server.mjs"] } } }
*Cursor, Windsurf & other MCP editors same stdio server

Any editor or agent that supports local MCP servers takes the same shape — point its MCP config at the built server with node as the command:

{ "command": "node", "args": ["/path/to/simple-web-crawler-scraper-extension/mcp-server/dist/server.mjs"] }

One extension, one browser session, several MCP clients: the first process to start becomes the hub and the rest relay through it automatically — run Scrapeline from two tools at once without a conflict.

☁claude.ai (browser) needs a tunnel — read this first

claude.ai runs entirely in the cloud, so it has no path to a process on your machine — a local stdio server is invisible to it by construction, not by missing config. Reaching it from there means putting a public URL in front of the local server (ngrok, Cloudflare Tunnel) with real authentication first, since Scrapeline currently trusts whatever calls it and reads your live, logged-in tabs. For most people, Claude Code or Claude Desktop gets the same capability with none of that exposure.

SAFETY MODEL

Local by construction, not by promise.

127.0.0.1 only

The bridge refuses any connection that isn't from your own extension's origin — a web page cannot reach it, and neither can anything off-machine.

Kill switch, always visible

One toggle in the popup turns Claude's access off entirely — no restart needed, and it holds until you turn it back on.

Content-free audit log

Every call is logged — method and host only. What was actually read is never written to disk.

Page text is untrusted data

Everything read from a page is labelled untrusted before it reaches Claude — instructions hidden in a page are never treated as commands.

No general clicking

The only clicks anywhere are the crawler's own pagination controls — restricted to things that look like "Next," and refused outright on anything labelled destructive.

Polite by default

Crawls wait between pages, check robots.txt, and cap themselves on pages, rows, and time — every job, not just the ones you remember to configure.

Two minutes to a connected browser.

No account, no cloud step, no billing page. Clone it, build it, load it.

git clone https://github.com/surypratapsingh/simple-web-crawler-scraper-extension.git && cd simple-web-crawler-scraper-extension npm install && npm run build

Then load extension/.output/chrome-mv3 as an unpacked extension, and point Claude Code, Claude Desktop, or another MCP client at mcp-server/dist/server.mjs. Full walkthrough in the project's README.