We've been hearing a lot about AI agents lately, and I've been putting many of them to the test. The excitement really began a few years back with tools like AutoGPT and CrewAI. While their approaches differ, they generally involve LLMs communicating with each other and using external tools. These "tools" are essentially APIs that bridge the gap between the LLM and the real world. An LLM alone can generate impressive text, but it's far more interesting when it can interact with its environment.
I recently explored Claude Computer Use, a new technology from Anthropic that lets Claude control your computer through natural language, with the assistance of some additional software. In theory, you could tell Claude, with this software running in a Docker container, to book a flight for you.
I emphasize "in theory" because we're still in the early stages, and things aren't always smooth. A major hurdle is speed: current methods often rely on taking screenshots, and then using the mouse to click and type. Even on a good computer, this can be slow, although it may still save you time in the end.
Despite these challenges, the technology is rapidly advancing and holds the potential to revolutionize numerous industries—a point underscored by the exciting new development I'm about to introduce.
---
What Is This Blog Post About?
Now, Claude is taking another giant leap towards truly agentic behavior with the introduction of the Model Context Protocol (MCP). You can dive into the details on the Claude website and GitHub, but simply put, MCP is a protocol that lets Claude connect to various applications and API on your computer or anywhere else.
All it takes is a simple configuration that you copy and paste into Claude.
Suddenly, Claude can interact with your file system, Google Drive, or even tools like Puppeteer for browser automation.
Here is a list of MCP Servers.
https://github.com/modelcontextprotocol/servers
---
How Does It Help Agentic Behavior?
On its own, Claude is limited to its internal knowledge, lacking access to the everyday services and applications you rely on, such as Google Drive, Gmail, or your computer's file system. Without this connection, Claude's capabilities are restricted to conversation—interpreting information and generating text.
However, with MCP configuration, you can integrate these tools, giving Claude the ability to interact with your broader digital ecosystem and perform real-world tasks for you.
---
How Does This Differ From LLM Tools and Function Calls?
The key difference lies in its runtime integration: you can add functionality to Claude on the fly with a simple configuration. Claude (Desktop Edition) then automatically handles the server connection, instantly gaining access to the MCP server's capabilities, essentially its API.
For instance, if the MCP server provides access to a file system, Claude receives a list of supported functions for reading and writing files and directories. It's important to understand that Claude only interacts with the API; the MCP server handles the actual operations on the file system.
You can examine the internal workings of an MCP file system server in the following code:
https://github.com/modelcontextprotocol/servers/blob/main/src/filesystem/index.ts
How does MCP differ from Custom GPTs?
Top comments (0)