MCP server

Give your agents the web with one MCP server

Add https://mcp.grabber.sh to your MCP client and your agent gets a fetch_page tool that reads any public page as clean Markdown, paid from the same prepaid credit as the API.

Get API key API docs

Remote: nothing to install
Claude, Cursor, VS Code, Windsurf
Same key and credit as the API

Set it up in a minute

Paste the config for your client and replace the key with your own.

Create an API key in the dashboard: dashboard.grabber.sh/keys
Add Grabber to your client with the config on the right
Restart the client and ask it to read a page

~/.cursor/mcp.json

{
  "mcpServers": {
    "grabber": {
      "url": "https://mcp.grabber.sh",
      "headers": {
        "Authorization": "Bearer gsh_your_api_key"
      }
    }
  }
}

Server details

Setting	Value
URL	`https://mcp.grabber.sh`
Transport	Streamable HTTP
Authentication	`Authorization: Bearer gsh_…`, the same key as the API
Tools	`fetch_page`, `check_balance`
Pricing	Same as the API: $0.001 a page, $0.005 rendered, errors free

Tools

Two tools, so the agent always knows which one to call.

Tool	What it does	Arguments	Cost
`fetch_page`	Loads one public web page and returns its content, as Markdown unless you ask for another format.	`url` (required): The http or https address to load. `format`: `markdown`, `html`, `text`, `links`. Default `markdown`. `js_render`: Run the page in a headless browser first. Default `false`. `max_chars`: Cut the content after this many characters, at most 100,000.	$0.001, or $0.005 with `js_render`. Errors are free.
`check_balance`	Returns the credit left on the account of the key, in US dollars.	None	Free

What agents do with it

Code against current docs

Point Cursor or Claude Code at a library's docs and it writes code for the version you use, not the one in its training data.

Research with sources

Ask for a summary of a page, a comparison of several, or a check of a claim against the original text.

Read single-page apps

The agent can set js_render when a page is built by JavaScript, so dashboards and app pages come through.

Keep the context small

Markdown is a fraction of the size of the HTML, and max_chars caps every page at 100,000 characters or less.

Try these prompts

"Read https://example.com/pricing and put the plans in a table."
"Fetch the docs page for the library we use and update our code to its current API."
"Read these three articles and tell me where they disagree."
"How much Grabber credit do I have left?"

MCP questions

Something else? The docs cover every parameter and error, and you can reach us from the dashboard.

Do I need to install anything?

No. The server is remote, at mcp.grabber.sh. Clients that support remote servers connect to it directly. Claude Desktop connects through mcp-remote, a small bridge that npx runs for you.

What does the MCP server cost?

The same as the API, from the same prepaid credit: fetch_page costs $0.001 a page, or $0.005 with js_render. check_balance is free, and so is any request that fails.

Why is a page capped at 100,000 characters?

So one long page cannot fill the context window of your agent. Set max_chars lower to save more room. The price is the same whatever the cap.

Is my API key safe in the config file?

The key stays in a file on your computer and is only sent to mcp.grabber.sh, over HTTPS. Use a separate key for each client, so you can see its usage and delete it on its own.

Which clients work?

Any client that can connect to a remote MCP server over Streamable HTTP and send a header, such as Cursor, Claude Code, VS Code and Windsurf. Claude Desktop works through mcp-remote.

Add Grabber to your agent

Get API key Read the docs