Build Your Own AI Assistant with Claude and MCP

What if Claude could read files on your computer. Search the web for you. Connect to your own systems.

It can. And it's easier to set up than you think.

This guide shows you how to build an AI assistant that can actually do things, not just answer questions. We'll use Claude Desktop and something called MCP (Model Context Protocol).

What We're Building

By the end of this guide, you'll have a Claude assistant that can:

• Read and write files on your computer
• Search the web for you
• Understand context from your local projects

You can then build on this with more tools as needed.

Prerequisites

You'll Need

1. Claude Desktop (Mac or Windows)
2. Node.js installed (version 18 or later)
3. A text editor (VS Code, or something else)
4. A bit of courage to use the terminal

Install Claude Desktop

If you don't have it already:

1. Go to claude.com/download
2. Download for your operating system
3. Install and log in with your Anthropic account

Install Node.js

If you don't have Node.js:

1. Go to nodejs.org
2. Download the LTS version
3. Run the installer

Verify it works by opening the terminal and typing:

node --version

You should see a version number.

Step 1: Find the Configuration File

Claude Desktop uses a JSON file to configure MCP servers. First we need to find it.

On Mac

The configuration file is at:

~/Library/Application Support/Claude/claude_desktop_config.json

You can open the folder by typing in the terminal:

open ~/Library/Application\ Support/Claude/

On Windows

The configuration file is at:

%APPDATA%\Claude\claude_desktop_config.json

Open File Explorer and paste that path.

Step 2: Install Your First MCP Server (Filesystem)

We'll start with the filesystem server. It lets Claude read and write files in folders you choose.

Install the Server

You actually don't need to install the server separately. With the npx configuration in the next step, it's fetched automatically the first time Claude starts. If you'd still rather pre-install it, run:

npm install -g @modelcontextprotocol/server-filesystem

Configure Claude

Open claude_desktop_config.json in your text editor and add:

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-filesystem",
        "/Users/yourname/Documents",
        "/Users/yourname/Projects"
      ]
    }
  }
}

Replace /Users/yourname/Documents and /Users/yourname/Projects with the folders you want Claude to access.

Restart Claude Desktop

Close Claude Desktop completely and open it again. MCP servers load at startup.

Test

Type something like:

Can you list the files in my Documents folder?

If everything works, Claude should be able to see and describe the contents.

Step 3: Add Web Search

Claude Desktop now has built-in web search, so for simpler needs you can get by without a separate search server. If you want more control, with your own quota and structured results, you add a search MCP. We'll use Tavily, which has a free tier of 1,000 API credits per month (one search uses at least one credit).

Get an API Key

1. Go to tavily.com
2. Create an account
3. Get an API key (the free tier covers 1,000 API credits/month)

Install and Configure

Update your claude_desktop_config.json:

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-filesystem",
        "/Users/yourname/Documents",
        "/Users/yourname/Projects"
      ]
    },
    "tavily": {
      "command": "npx",
      "args": [
        "-y",
        "tavily-mcp@latest"
      ],
      "env": {
        "TAVILY_API_KEY": "tvly-YOUR-API-KEY"
      }
    }
  }
}

Replace tvly-YOUR-API-KEY with your actual API key.

Restart and Test

Restart Claude Desktop. Test with:

Search the web for the latest news about AI agents

Claude should now be able to search and summarize results.

Step 4: Combine the Tools

Now you have an assistant that can both read local files and search the web. Try combining:

I have a project in /Users/yourname/Projects/my-app.
Can you read the README file and then search the web for
best practices for the framework the project uses?

Or:

Summarize the contents of all markdown files in my
Documents folder and create a new file with the summary.

More MCP Servers to Explore

Filesystem and search are just the beginning. Here are other MCP servers that might be interesting. One thing to know: the early reference servers for GitHub, Slack, Postgres and others are now archived and no longer maintained. So below I point to the actively maintained alternatives.

GitHub

Let Claude read and edit code in your repositories. Use GitHub's own official server, github/github-mcp-server. The easiest path is to run it via Docker (ghcr.io/github/github-mcp-server) or connect the hosted remote version directly. The repo's README has ready-made configs for Claude.

Slack

Let Claude read and send messages. The most-used community server is korotovsky/slack-mcp-server, which has write actions disabled by default.

PostgreSQL

Let Claude query your database. crystaldba/postgres-mcp (Postgres MCP Pro) gives configurable read/write access plus performance analysis, and installs via Docker, pipx or uv.

Find more actively maintained servers in the official MCP registry and the MCP servers repo.

Building Custom MCP Servers

If you have a technical background, you can build custom MCP servers to connect Claude to exactly the systems you want. It requires some coding but opens up unlimited possibilities.

MCP uses standardized JSON-RPC based communication. There are SDKs for Python and TypeScript that make it relatively easy to get started.

That's outside the scope of this guide, but the tip is to start by studying existing MCP servers as examples.

Security: What to Think About

Before giving Claude access to more tools:

Permissions

Only give access to what's needed. Avoid broad permissions like "the entire filesystem."

Sensitive Data

Be careful with databases and systems containing customer data or trade secrets.

Review Output

Especially in the beginning, check what Claude actually does. If it's going to write files or send messages, verify it looks right.

API Keys

Store them securely. They give access to your accounts and services.

Troubleshooting

Claude Doesn't See the MCP Servers

1. Check that the JSON syntax in the config file is correct
2. Restart Claude Desktop completely (not just minimize)
3. Verify Node.js is installed and working

Filesystem Server Can't Find Files

1. Double-check the paths in the config file
2. Make sure the folders actually exist
3. Check permissions (especially on Mac)

Web Search Doesn't Work

1. Verify the API key is correct
2. Check you haven't exceeded your monthly quota
3. Test the key directly against Tavily's API to isolate the problem

TLDR

1. MCP gives Claude the ability to use tools and do things, not just respond.
2. Filesystem server lets Claude read and write local files.
3. Tavily provides web search (free tier of 1,000 API credits/month), or use Claude Desktop's built-in search.
4. Combine tools for powerful workflows.
5. Think about security and only give access to what's needed.

The nice thing about this setup is you can start simple and build out gradually. Add new MCP servers as you find new needs.

Want to see how I use MCP and Claude in my daily work? Read AI for consultants: how I use it in my own work.

Frequently Asked Questions

Want help setting up a customized AI assistant for your business? Learn more about our AI development service or book a call and we'll discuss what's possible.