If you've been using Claude, Cursor, or Windsurf lately, you may have seen the term "MCP server" pop up. People are building them, publishing them to directories, and talking about them as if they're the next app store.
Here's what they actually are — explained without jargon.
The problem MCP solves
AI assistants are good at reasoning. But they're isolated. Out of the box, Claude can't check your calendar, scan a receipt, look up a stock price, or run a query against your database. It can only work with what you paste into the chat.
There have always been workarounds — plugins, function calling, custom integrations — but each one was bespoke. Every AI tool had its own way of connecting to external services, and every developer had to implement it differently for every client.
MCP is an attempt to standardize that. One protocol. Any AI client. Any tool server.
The USB-C analogy
Think of MCP like USB-C for AI tools.
Before USB-C, every device had its own charging port. You needed a different cable for your phone, your laptop, your headphones. Annoying.
USB-C created a standard: one port, works everywhere.
MCP does the same thing for AI tools. Before MCP, if you wanted Claude to access your receipts, you'd build one integration. For ChatGPT, you'd build a different one. For Cursor, another one. Same capability, three separate implementations.
With MCP, you build one server. Any MCP-compatible client — Claude, Cursor, Windsurf, any future client — can connect to it without modification.
How it actually works (simply)
An MCP server is a small program that:
- Declares tools — it tells the AI client "here's a tool called
convert_receipt, here's what it does, here's what inputs it takes" - Handles calls — when the AI decides to use the tool, the server executes it and returns the result
- Runs locally — it runs on your machine, not in the cloud
The AI client (Claude, Cursor, etc.) handles the rest: deciding when to call the tool, passing the right inputs, presenting the results.
Here's a concrete example. When you connect the receiptconverter-mcp server to Claude Desktop:
You: Parse this receipt for me: https://example.com/lunch.jpg
Claude: [internally decides to call convert_receipt] [MCP server sends the URL to ReceiptConverter API] [gets back JSON: vendor, total, line items, taxes]
Claude: Here's what I found:
- Vendor: Chipotle
- Total: $14.75
- Items: Burrito Bowl ($12.00), Chips & Guac ($2.75)
- Category: Food & Drink
You never wrote any code. Claude never needed to know the API endpoint, the auth format, or the response schema. The MCP server handled all of that.
What MCP servers exist right now
The ecosystem is growing fast. As of early 2026, there are MCP servers for:
- Developer tools: GitHub, GitLab, Jira, Linear
- Productivity: Notion, Google Drive, Slack, Todoist
- Finance: Stripe, QuickBooks, ReceiptConverter
- Data: PostgreSQL, SQLite, Airtable, Google Sheets
- Utilities: web search, browser automation, file system access
You can find them at mcpmarket.com, glama.ai/mcp/servers, and the Anthropic MCP directory.
Who builds MCP servers?
Both sides of the equation:
SaaS companies build MCP servers to make their APIs accessible to AI tools. ReceiptConverter has one (receiptconverter-mcp). Stripe, GitHub, and Notion have them too.
Developers build custom MCP servers to connect their own internal tools — a company database, a proprietary API, an internal knowledge base.
Do you need to be a developer to use MCP?
To use an existing MCP server: no, not really. If you can edit a JSON config file and restart an app, you can install an MCP server.
Here's how simple the config is for Claude Desktop:
{
"mcpServers": {
"receiptconverter": {
"command": "npx",
"args": ["receiptconverter-mcp"],
"env": {
"RECEIPTCONVERTER_API_KEY": "sk_live_your_key"
}
}
}
}
That's it. Edit the file, restart Claude, done.
To build an MCP server: you need basic Node.js or Python skills. The core of a minimal server is about 50 lines.
Why it matters
MCP matters because it shifts how people interact with software.
Right now, you install apps. You log in to apps. You switch between apps to do things.
The MCP future looks different: you interact with an AI that has tools. You say "scan this receipt and add it to my expense spreadsheet" — and the AI calls convert_receipt, gets the data, calls the Sheets tool, writes the row. You never opened a tab.
That's not science fiction. It's working today in Claude Desktop, Cursor, and Windsurf.
Try it with ReceiptConverter: docs/mcp — setup takes 2 minutes, Node.js required.