GuideModel Context ProtocolIntermediate

Connect Claude to your work tools with MCP, safely

What the Model Context Protocol is, how to add an MCP server to Claude as a connector or a local server, and what to check before you let it act on your behalf.

Curated by Javier Díaz · Written with Claude from official documentation · Checked 18 October 2026 · 25 min

The Model Context Protocol (MCP) is an open standard that lets an AI app talk to outside tools and data in a common way. A service publishes an MCP server once, and any app that speaks MCP can use it, so you are not waiting for each vendor to build its own integration. Anthropic created MCP and in December 2025 donated it to the Agentic AI Foundation, a fund run under the Linux Foundation. The project is now set up as an LF Projects entity and is run by its own group of individual maintainers.

This guide is for professionals who already use Claude and want it to read from, or act in, the tools they work with every day: a project tracker, a document store, a code host, a folder on their laptop. You will connect one server, test it on a harmless task and learn which permissions to keep tight.

Use MCP when you keep copying information between Claude and another app, or when you want Claude to look something up for itself rather than rely on what you paste in. If a ready-made connector already exists in Claude's directory, start there. The custom route is for services that publish their own MCP server.

1. 01Learn the three roles: host, client and server

   MCP describes a conversation between three parties. It helps to know which is which before you start changing settings.

   • Host: the AI app you actually use, such as Claude Desktop, Claude on the web or VS Code. It decides what the model is allowed to do.
   • Client: a small connection the host opens for each server. You rarely see it; it's plumbing.
   • Server: the program that offers capabilities. It can run on your machine (a local server, which talks to the host over standard input and output) or on the internet (a remote server, reached over HTTP).

   Servers offer three kinds of thing: tools the model can call to do something, resources that supply data to read, and prompts, which are ready-made templates. Most of what you will notice day to day is tools.

2. 02Pick one server for one job

   Start with a single, specific need, for example "find the open tickets assigned to me" or "summarise the documents in this folder". Then look for a server that does that job.

   • Check Claude's Connectors Directory first (claude.ai/connectors). Each listing says what the connector does, whether it reads or writes, and where it works.
   • If your service isn't listed, see whether the vendor publishes its own remote MCP server and gives you its URL.
   • Prefer servers built and hosted by the company that owns the data. A third-party wrapper around someone else's API adds a party you have to trust.

   Resist connecting five servers at once. Each one adds tools the model can call, and problems are much easier to trace when there is only one.

3. 03Check the server before you connect it

   Anthropic says plainly that it does not vet custom connectors, so this step is yours. Before you add anything, answer these questions:

   • Who runs it? A named company with a security page is a better bet than an anonymous repository.
   • What can it do? Read the tool list. Can it only read data, or can it also create, change or delete things and send messages as you?
   • What will it ask for? When you sign in, the service will show the scopes it wants. Grant the narrowest set that does the job and decline the rest.
   • Does it use OAuth? That means Claude never sees your password and you can revoke access later. A shared API key is cruder: on Team and Enterprise everyone using the connector gets that key's full access.
   • Can it reach Claude? Remote connectors are called from Anthropic's cloud, not your device, so a server that only lives on your company VPN will not connect unless your network team allowlists Anthropic's IP ranges.

   For local servers the stakes are higher: they run on your computer with your user account's permissions.

4. 04Add a remote server as a custom connector

   On Pro or Max, you add the connector yourself:

   1. In Claude, open Customize, then Connectors.
   2. Choose to add a custom connector and give it a name you will recognise.
   3. Paste the server's URL (it should start with https://) and continue.
   4. Claude detects how the server expects you to sign in. Accept the defaults unless the server's documentation tells you otherwise. If it uses fixed credentials such as an API key, there is a place to add request headers.
   5. Finish, then sign in to the service when prompted and review the permissions it requests.

   On Team and Enterprise the work is split. An Owner adds the connector in Organization settings under Connectors, and then each member finds it in their own Connectors list, usually marked Custom, and signs in individually. Claude then sees only what that person could already see in the service.

   If you just want to try the mechanics first, the MCP documentation offers a public demo server at https://example-server.modelcontextprotocol.io/mcp. It holds no data of yours.

5. 05Switch it on per chat and decide what it may do unasked

   Adding a connector does not mean every conversation uses it. In a chat, open the + menu at the bottom left, go to Connectors and toggle on only the ones that conversation needs.

   When Claude wants to call a tool it asks for approval, and you can choose to always allow it. Be choosy:

   • Allow read-only tools freely once you're happy with what they return.
   • Keep approval on for anything that writes, deletes, sends or pays.
   • Turn off tools that are irrelevant to the task, so the model is not tempted to use them.

   Team and Enterprise Owners can go further and set each connector's tools to always allowed, needs approval or blocked for the whole organisation.

6. 06Optional: run a local server in Claude Desktop

   Local servers suit things that live on your computer, such as a folder of files. They work only in Claude Desktop, not in Claude on the web or Cowork.

   The simple route is a desktop extension: a packaged local server you install from Settings > Extensions in Claude Desktop, entering any required keys in a form. Claude Desktop ships with its own Node.js for these.

   The manual route is a JSON file. In Claude Desktop's own Settings window, open the Developer tab and choose Edit Config. This opens claude_desktop_config.json. A minimal entry that gives a file server access to a single sandbox folder looks like this:

   {
     "mcpServers": {
       "sandbox-files": {
         "command": "npx",
         "args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/you/Documents/claude-sandbox"]
       }
     }
   }
   

   Use an absolute path, and on Windows write backslashes doubled (C:\\Users\\you\\...). Quit Claude Desktop fully and reopen it; the server should then appear under Connectors. Point it at a folder made for the purpose, never your whole home directory: the server can do anything to those files that you can.

7. 07Try a first task that only reads

   Your first real task should be useful but impossible to regret. Good choices:

   • "List the five most recently updated items in [project] and tell me which look blocked."
   • "Read the files in my sandbox folder and give me a one-paragraph summary of each."
   • "Find the last three documents that mention [client name] and list the decisions recorded in them."

   Watch the approval prompts as they appear. They show which tool Claude picked and what it sent, which is the quickest way to learn how the server behaves. Check the answer against the source app once. When you trust the reads, try a small write you can undo easily, such as adding a comment or creating a draft.

8. 08Review and prune your connectors

   Treat connectors like any other app you've given account access to.

   • Remove connectors you no longer use. In Claude this is done from the Connectors page; to change a connector's URL or credentials you remove it and add it again.
   • Revoke Claude's access from the third-party service's own security settings as well, if you're leaving it for good.
   • Pay attention when a server updates. If its tools start behaving differently, or ask for new scopes, stop and check before carrying on.
   • If you come across a server that looks malicious, Anthropic's help centre explains how to report it.

Written by Claude from the official pages above and checked against them on 18 October 2026. Products change; if a step no longer matches what you see, the sources have the current version. How we work