Introduction to mCP for Agent-API Communication

Introduction to mCP

• mCP is intended to unlock previously unattainable functionalities.
• It tackles complexities involved in agent-API communications.
• mCP is designed to simplify agent connection to APIs through a uniform protocol.

The Problem Without mCP

• Building agents that interact with APIs (like Slack, Gmail, or custom databases) involve creating integrations from scratch.
• APIs like Gmail might allow message deletion, but restrictions may be needed for agents to only create or save messages.
• Custom prompts or implementations become necessary to restrict agents to particular actions within an API.
• Custom implementations for one agent (e.g., interacting with Slack and Gmail in WindServe) aren't transferable to other agents like Cursor.
• Custom work must be re-implemented for each new agent.

mCP as a Solution

• mCP reduces complexity in agent-API interactions.
• It enables the sharing of connections, implementations, and specifications across different agents.
• A standardized protocol is used for LLMs or agents when connecting to backends or APIs.
• Agents, including custom agents, Clock Desktop, Cursor, or WindServe, can use the protocol to communicate with an mCP server.
• The mCP server can manage connections to Slack, Gmail, and custom databases.
• Custom specifications are shifted to the mCP server, which eliminates the need for agent-specific reimplementation.

How mCP Works

• An mCP server acts as an intermediary, managing connections to various services/APIs.
• Client applications such as IDEs can establish connections with company-provided mCP servers (e.g., Slack's mCP server).
• This provides universal access to services from different agents, such as Cursor accessing Slack.
• Additional information is available on the mCP website, under Core Concepts.

Core Concepts of mCP

• Host: The application that implements the agent (e.g., Cursor).
• Client: A tool that communicates with the server.
• Server: Contains the implementation to be accessed (e.g., Slack's mCP server, Gmail's mCP server, or a custom server).

Functionalities Within mCP Servers

• Tools: Functionalities like a calculator or weather service are exposed from the mCP server.
• A client application, like Cursor, can connect to the mCP server and list available tools.
• Tools function similarly to tools within Llama Index or Langchain.
• Prompts: Prompts for the LLM or agent to use can be shared.
• Resources: Documentation can be shared through the mCP server, which guides agents on API usage or other tasks.

Additional Capabilities of mCP

• mCP servers from different companies are independent (e.g., one for Slack, one for Gmail).
• Custom mCP servers have the capability to implement complex, simultaneous tasks.
• The mCP server can interact with Slack, Gmail, and a custom database to write logs.
• Agents (like Cursor) only need to communicate with the custom mCP server, which abstracts away the complexity of multiple interactions.
• Anthropic maintains a directory of mCP servers.
• Compos offers 250+ mCP servers available for plug-and-play integration.
• Numerous companies are exposing functionalities via mCP servers.

Demonstration: Using mCP with a Machine Learning Model

• Implementing an mCP server for testing a Machine Learning Model.
• The goal is to enable using cursor to test the model, replacing ad-hoc scripts.
• The mCP server has a tool called "invoke model" that can receive a payload.
• This tool sends POST requests to a specified URL with the received payload argument in JSON format.

Implementation of mCP Server

• A Python implementation simplifies communication between cursor and the model.
• Documentation is included in a Python sample on the mCP website.
• The demonstration involves invoking a model from within Cursor, based on data in a CSV file.

Configuring Cursor to Communicate with the mCP Server

• mCP servers can be added in Cursor's settings under features.
• A command is provided to run the mCP server which is a script on the local computer.
• Cursor connects to the mCP server to see available tools.
• This demonstrates the communication established between Cursor and the server upon connection.

Interacting with the Model Through Cursor and mCP

• Ensure that Cursor is in agent mode to enable mCP tooling
• Cursor locates, reads the "penguins.csv" file, and understands the data format.
• It uses mCP to send samples from the dataset to the model.
• Cursor formats the request and sends it to the "invoke model" tool on the server.
• The model is invoked with the sample data, and a prediction is obtained successfully.

Experimenting with the Model Through Cursor

• Instructions are given to cursor to change the body weight, resend the data and invoke the model again.
• Cursor calculates the mean body mass and standard deviation, then sends samples with updated mass.
• Cursor communicates the model's reaction based on the weight change.
• Cursor calculates an average derived from the model, facilitated by the free interaction enabled through AI implementation.

Advantages

• The implemented MCP server can be used across other agents like WindServe or Cloud Desktop.
• Communication and interactions are facilitated without repetitive code development.