DEV Community

Cover image for Give Claude or Cursor the web: MeshArc MCP in 60 seconds
Sid Mishra
Sid Mishra

Posted on Originally published at mesharc.dev

Give Claude or Cursor the web: MeshArc MCP in 60 seconds

I kept asking my coding assistant about libraries released last month and competitor pricing pages as they stand today, and kept getting confident guesses from memory. So I built MeshArc a web scraping MCP server: it hands the assistant tools that fetch a page as clean Markdown, list every URL a site declares, crawl a section, and report what changed since the last visit.

This guide connects MeshArc's hosted MCP server to Claude, Claude Code and Cursor. The hosted server has no package to install and no key to copy. Then it shows one real run, with the numbers as they came back: two tool calls, 2 credits.

What you need

  • A MeshArc account. The free plan gives 1,000 credits once, with no card.
  • Claude (web, desktop or mobile), Claude Code, Cursor, or any client that accepts a remote MCP URL.
  • One URL: https://mcp.mesharc.dev/mcp

That is the whole list for the hosted server. If you would rather run the server on your own machine with your own API key, skip to "Run it locally instead".

Step 1: add the server

Claude (web, desktop and mobile)

  1. Open Settings → Connectors and choose Add custom connector.
  2. Paste https://mcp.mesharc.dev/mcp and give it a name.
  3. Sign in to mesharc.dev when the browser asks, and approve.

Claude Code

One line in a terminal:

claude mcp add --transport http mesharc https://mcp.mesharc.dev/mcp
Enter fullscreen mode Exit fullscreen mode

It opens a browser once to sign in and approve, then works.

Cursor

Use the Add to Cursor button on the MCP page, or add the server to ~/.cursor/mcp.json yourself (.cursor/mcp.json inside a repository keeps it to that project):

{
  "mcpServers": {
    "mesharc": {
      "url": "https://mcp.mesharc.dev/mcp"
    }
  }
}
Enter fullscreen mode Exit fullscreen mode

Cursor asks you to confirm, then you sign in to mesharc.dev once.

VS Code and everything else

The MCP page also has an Add to VS Code button, which registers the server as {"type": "http", "url": "https://mcp.mesharc.dev/mcp"}. ChatGPT connectors and any other client that takes a remote MCP URL use the same address.

Step 2: approve, and tick the right box

The approval screen matters more than it looks. Approving does not hand the assistant your API key. MeshArc issues the connected app a key of its own, which you can see and revoke under Settings → API keys. The assistant never holds a credential.

A connection is read only unless you allow changes, and read only means less than most people expect. A read-only connection can look at everything already stored in your workspace (projects, runs, pages, change records), but it cannot reach a website. Anything that fetches counts as a write, because it leaves the workspace and usually spends credits. That includes reading a single page and mapping a sitemap.

So if the goal is to give Claude or Cursor the web, tick Also allow changes when you approve. If you only want the assistant to query crawls you have already run, leave it off. A tool called without the right access answers {"code": "read_only"} and says what to do about it; nothing fails silently.

Step 3: ask for a page

Now ask in plain English. Two prompts are enough to see the two tools you will use most.

"How many pages does the FastAPI docs site declare, and which are in the security tutorial?" The assistant calls map_site, which reads a site's sitemaps without fetching any page.

"Read the first one and summarise it." The assistant calls scrape_urls, which returns the page as Markdown.

One real run, with numbers

We ran exactly that through the hosted server on 12 November 2026. Results as they came back:

Call What came back Credits
map_site on fastapi.tiangolo.com, filtered to /tutorial/security/ 153 URLs declared in 1 sitemap file; 5 under /tutorial/security/ 1
scrape_urls on /tutorial/security/first-steps/ 2,291 words; 127,298 bytes of HTML reduced to 10,400 bytes of Markdown; plain fetch, verdict ok 1
Total The size of the site, the section, and one page read 2

Two details in that table are worth a second look. The map cost 1 credit because it is charged per sitemap file, not per URL, so learning the size of a 153-page site costs the same as learning the size of a 15,000-page one that fits in one file. And the Markdown is about 8% of the HTML by size. That matters inside a chat: the assistant's context holds the content of the page, not its navigation, sponsor banners and script tags.

The same two calls from a script, with the Python SDK:

pip install mesharc
Enter fullscreen mode Exit fullscreen mode
from mesharc import MeshArc

arc = MeshArc()  # reads MESHARC_API_KEY

details = arc.map_details("https://fastapi.tiangolo.com/", search="/tutorial/security/")
print(details["totals"], details["creditsUsed"])

page = arc.scrape("https://fastapi.tiangolo.com/tutorial/security/first-steps/")
print(page["verdict"], page["credits"])
print(page["markdown"][:500])
Enter fullscreen mode Exit fullscreen mode

A tool call through MCP is an ordinary authenticated request, so the credits, limits and rate are identical either way.

Beyond one page: watch a site

Reading a page is where most web MCP servers stop. The larger group of MeshArc's tools is for a site you come back to:

  • crawl_site crawls a site once, to a limit and depth, with include and exclude globs.
  • create_project turns a seed URL into a watched project with a schedule: manual, hourly, daily or weekly.
  • start_run crawls it now, or only the pages you list.
  • get_changes lists the pages added, modified and removed since the run before.
  • list_pages searches the stored pages by word, quoted phrase or CSS selector.

The MCP page shows a conversation of this kind as it ran. Asked to "set up a project that only scrapes the blog, weekly", the assistant first called get_project with no project, which returns every setting it may set. It then mapped the site, found the blog under /blogs/ and not /blog/, and created the project with include_paths set to ["/blogs/*"]. The first run read 30 blog pages and nothing else.

Setting names are exact. A guessed key such as includePaths is refused with its name, so the assistant corrects itself and does not go on believing a project is configured when it is not. The full list is on the configuration page.

Reading what is stored is free. list_projects, list_runs, list_pages, get_changes and get_job cost 0 credits, so "what changed on the docs since Monday?" costs nothing to ask as often as you like. If the pages are feeding a retrieval index, how to keep a RAG knowledge base fresh covers the other half, and the RAG pipeline use case shows the whole route from crawl to vector store.

What it costs

There is no charge for the MCP server itself. You pay for the work, at the same rates as the API (pricing page, checked 11 November 2026):

Work Credits
A page read by plain fetch 1
A page that needs a real browser 4
A sitemap file read by map_site 1
A page the site refused, a 404 or a 410 0
Listing projects, runs, pages and changes 0

On the Growth plan ($99 a month, 500,000 credits) a plain page works out at $0.20 per 1,000. The free plan's 1,000 credits cover a lot of chat-sized questions: the run above used 2. Every response states what it cost. Full rates are on pricing.

Run it locally instead

Some teams do not connect hosted apps, or want the server to hold a narrowly scoped key. The server ships in the Python package (Python 3.10 or newer) and speaks stdio:

pip install "mesharc[mcp]"
claude mcp add mesharc -e MESHARC_API_KEY=mesharc_... -- mesharc-mcp
Enter fullscreen mode Exit fullscreen mode

For Claude Desktop or Cursor, the same thing as a config entry:

{
  "mcpServers": {
    "mesharc": {
      "command": "mesharc-mcp",
      "env": { "MESHARC_API_KEY": "mesharc_..." }
    }
  }
}
Enter fullscreen mode Exit fullscreen mode

Make the key under Settings → API keys. A read key lets the assistant look at what is stored; write lets it fetch and run projects. A key can also be limited to named projects, and given an expiry. One difference from hosted: deleting a project needs an admin key, which a connected app never holds, so delete_project only works locally.

What it will not do

Better to know before the assistant finds out mid-conversation.

Limit What happens
Pages behind a sign-in The crawler reads what a visitor can see. A login wall comes back as blocked, at 0 credits
Checking a page every minute Schedules are hourly at the finest. start_run with the pages listed is the on-demand check
A 5,000-page crawl inside one chat turn Hosted, a tool that would hold the connection open hands back a job after 45 seconds; get_job picks it up. Start a big site as a project and read it back with list_pages
A very long page in full One page asked for on its own is capped at 12,000 characters in the tool's answer. The API returns the whole page
A read-only connection that fetches Refused with read_only until you reconnect with changes allowed
More than the plan allows The same credits and limits as a script, and the same 402 when they run out

If all you need is one page fetched on demand and never revisited, a single-purpose fetch tool is less to set up, and your assistant may already have one built in. This server earns its place when the assistant should remember a site and tell you what moved.

The short version

  • Add https://mcp.mesharc.dev/mcp as a connector in Claude, with claude mcp add --transport http in Claude Code, or in mcp.json in Cursor.
  • Sign in and approve. Hosted server: no key to copy. Tick Also allow changes if the assistant should fetch pages.
  • map_site sizes a site for 1 credit per sitemap file; scrape_urls reads a page for 1 credit by plain fetch or 4 with a browser.
  • Projects, runs and get_changes turn "read this page" into "tell me what changed".
  • Refused pages, 404s and reads of stored data cost nothing.

Try a URL in the playground to see what the assistant will get back from it, or start free with 1,000 credits and no card, then paste the server URL into your assistant.


Originally published on mesharc.dev.

Top comments (0)