MCP: the USB-C of AI tools
Model Context Protocol — the standard that lets any AI model plug into any tool, and why that's more important than any single model.
The problem before MCP
Every AI tool needed its own way to connect to the outside world:
- Tool A → custom plugin format → 50 integrations
- Tool B → different plugin format → another 50
- Tool C → REST API wrapper → another 30
The result: N tools × M integrations = N×M connection problems. Every new model, every new tool, every new data source was a new integration project. And nothing was portable — the integration you built for Claude didn’t work in ChatGPT, and vice versa.
The MCP fix
Model Context Protocol (MCP) is a standard — like USB-C, like HTTP, like USB-C for AI. One protocol, one way to connect:
┌─────────────┐ ┌──────────────────┐ ┌─────────────┐
│ AI model │◄───────►│ MCP server │◄───────►│ Your tool │
│ (any LLM) │ MCP │ (the bridge) │ MCP │ (calendar, │
│ │ │ │ │ files, DB) │
└─────────────┘ └──────────────────┘ └─────────────┘
The MCP server wraps your tool (a calendar, a database, a file system) and exposes it over a standard protocol. Any MCP-compatible client — Claude Desktop, VS Code, your custom app — can plug in without writing a single line of integration code.
| Before MCP | With MCP |
|---|---|
| 1 integration per tool × model | 1 integration per tool, works with all models |
| Vendor-locked plugins | Open standard, cross-platform |
| “Can I use this with my AI?” | “Does it speak MCP?” |
What an MCP server exposes
An MCP server typically offers three kinds of capabilities:
- Tools — functions the model can call (search a database, send an email, read a file)
- Resources — data the model can read (a document, a config file, a log)
- Prompts — pre-built prompt templates (a “code review” prompt, a “summarize” prompt)
The model sees the names and descriptions of these tools. It decides when to call one. The MCP server executes it and returns the result. That’s the loop — and it’s exactly the agent loop (think → act → observe) that the agents section explains in detail.
Why this matters more than any single model
Models get better every quarter. The tool ecosystem is the durable part. A well-built MCP server for your internal knowledge base works today with the current model and will work with next year’s model — without you rewriting a line.
That’s the same logic as why “supports USB-C” matters more than “has 8 GB RAM” when buying a laptop: the standard is the moat, the hardware is the churn.
The security note you can’t skip
MCP tools are execution surfaces. When a model calls an MCP tool, it’s running code with real side effects — writing files, sending emails, querying databases. The same prompt-injection attack that targets the model now has a path to your actual tools.
The practical rules:
- Least privilege. An MCP server for “read your calendar” should only be able to read your calendar. Not write, not delete, not read your emails.
- Audit your MCP servers the way you’d audit browser extensions. Every server is a plugin with real access.
- Log tool calls. If something unexpected happens, you need to know which tool was called, with what arguments, at what time.