Claude MCP servers for data work: how to add and vet them

You use Claude for real work now, and the work needs data it does not have. Search results, a list of companies, a competitor's keywords, a creator's recent posts. The answer in the docs is "add an MCP server". The answer you need is which ones, how many, and how to avoid a setup that grows into a mess.

This post covers Claude MCP servers for data work. It shows the command to add one, the checks to run on any server before you give it access, the reason fewer is better, and a prompt that tests the whole setup in one go.

What are Claude MCP servers?

An MCP server is a program or a hosted endpoint that gives Claude a set of tools over the Model Context Protocol. Claude lists the tools, decides when to use one, and calls it with the arguments it fills in. If the term is new, what is an MCP server explains the protocol in plain words.

For data work, the tools usually fall into five groups: web search, page scraping, SEO and keyword data, company and people enrichment, and social data. One server can cover one group or several.

How do I add an MCP server to Claude Code?

You add a server with a command or a config entry, then restart the session so Claude sees the tools. For a hosted server that speaks streamable HTTP, the command has this shape:

claude mcp add --transport http looot https://api.looot.ai/mcp

That adds the looot server. The looot CLI can also print a ready-to-paste config for your client with `looot mcp install --client claude-code`. Either way, the first use opens a browser sign-in. A headless agent uses a bearer token created in the dashboard instead.

Check the connection by asking Claude to list the tools it has from the new server. If the list is empty, the connection failed, and the fix is in the server's address or your sign-in, not in the prompt.

Which MCP servers does a data task need?

Start from the job, not from a directory of servers. Write down the three data jobs you do most, then find the smallest set of servers that covers them.

The providers behind those rows are in the public catalog, so one looot connection reaches them. Start with that one server and add a direct one only when you need a feature the catalog lacks.

How many MCP servers is too many?

More than you can name from memory is too many. Every server adds tool descriptions to Claude's context, and a long list makes the model slower and worse at choosing. The cost shows up as wrong tool picks, not as an error.

A simple rule works well. Keep one general server for data jobs and one or two for tools you own. If you find yourself adding a fourth data server, ask whether a gateway would replace three of them. The MCP gateway post explains when that makes sense.

What should I check before I trust an MCP server?

A server runs with the access you give it, so vet it like a dependency. Ask five questions, and write the answers down.

1. Who runs it? A hosted server means a company handles your calls and sees your inputs. A local server runs on your machine and sees what you let it read.
2. What can it do? List its tools. A read-only search tool is low risk. A tool that sends email or writes files is not.
3. What does it need? Note the tokens and scopes. Give the narrowest scope that works.
4. What does a call cost? Find out before the agent loops. A tool that bills per call needs a spend limit.
5. Can you see what ran? You want a log of calls, with inputs and costs, that you can read afterward.

For looot, the answers are as follows. A hosted gateway runs the calls. The agent can search and inspect for free and run endpoints that bill per call. The token has scopes for reading the catalog, reading runs and executing runs. Every endpoint shows its price before it runs, and `looot runs get` returns the cost of a finished call.

What does a data job look like end to end?

Take a small job: "find ten companies that sell project software to agencies and get a contact for each." Claude does not hold that data, so it uses the tools you connected.

First it calls a search tool for pages about agency project software and reads the results. Then it asks the catalog for a company search endpoint and a work email endpoint, and reads each price. It runs the search, picks ten companies, and calls the email endpoint for each name. At the end it writes a table with the company, the contact, the source endpoint and the cost.

Three habits make this reliable. Put the budget in the prompt. Ask for the source and cost on every row. And read the first three rows by hand before you accept the other seven. A model fills gaps with confident guesses, and the cheapest check is a person reading a few rows.

The Clay replacement post walks through a longer version of this job, with the places where an agent falls short.

Try it: a prompt that tests the setup

Run this after you add the server. It uses only free calls, so it tests the connection without spending anything.

If the answer to step 4 names a gap, add the next server for that gap and rerun the prompt.

What it cannot do

• A server does not make Claude correct. The model can still misread a result, so check any number you plan to publish.
• It does not give Claude access to your private data unless you connect it. A hosted data server sees only what the agent sends.
• It does not run work while Claude is idle. Anything that needs a timer needs scheduled runs, not available yet.
• It does not cap spend unless you set a cap. Tell the agent a limit in the prompt, and set a budget on the token where the server supports one.

Last checked 2026-10-03 against the looot skill file and the live catalog. Author: the looot team.