DIY MCP Part 3: AI Tool Calling

Key Concepts:

• DIY MCP (Message Communication Protocol) server and client
• AI integration (Anthropic Claude)
• Tool use and function calling in AI
• Message transcript for AI interaction
• MCP tool proxying to AI
• Input schema transformation for AI compatibility

1. Connecting to an AI Service

• The video focuses on connecting a DIY MCP client to an AI service, specifically Anthropic's Claude.
• Alternatives like OpenAI, O Lama, and Hugging Face are mentioned, but Anthropic is chosen for demonstration.
• The Versal AI library is recommended as a tool to abstract away the complexities of different AI systems' tool handling. However, the video avoids using it to demonstrate the underlying mechanisms.

2. The callAI Function

• A callAI function is created to interact with the Anthropic API.
• It takes messages (the chat transcript) and tools (tool definitions) as input.
• It uses fetch to make a POST request to the Anthropic API endpoint (v1/messages).
• The request includes the Anthropic API key, model (claude), max_tokens, messages, and tools.
• The function parses the JSON response and extracts the content.
• Example:

  async function callAI(messages, tools) {
    const response = await fetch("https://api.anthropic.com/v1/messages", {
      method: "POST",
      headers: {
        "X-API-Key": anthropicApiKey,
        "Content-Type": "application/json",
        "Anthropic-Version": "2023-10-01",
      },
      body: JSON.stringify({
        model: "claude-3-opus-20240229",
        max_tokens: 1024,
        messages: messages,
        tools: tools,
      }),
    });
    const data = await response.json();
    return data.content;
  }
  

3. Integrating AI into the MCP Client

• A new menu option "AI" is added to the MCP client.
• When the user selects "AI," they are prompted to enter a question.
• The question is added to the messages array.
• A callAIWithTools closure is created to handle the AI call with the available tools.
• The inputCamelCaseSchema in the tool definitions is transformed to input_schema to match Anthropic's required format.
• The callAI function is then invoked with the messages and transformed tools.

4. Handling Tool Use Requests

• The AI may respond with a "tool use" request, indicating that it needs to call a specific MCP tool.
• The code checks if the last item in the AI's response indicates a tool use request.
• If a tool use request is detected, the client extracts the tool name and arguments from the AI's response.
• It then makes an MCP call to the server, proxying the tool call request from the AI.
• The response from the MCP server (the tool result) is added to the messages array as a tool_result.
• The tool_use_id from the AI's request is linked to the tool_result to track the call.
• The callAI function is called again with the updated messages array, allowing the AI to process the tool result and provide a final answer.

5. Example Scenario: Coffee Shop

• The example uses a coffee shop scenario with MCP tools for getting drink names and drink information.
• The user asks the AI "What drinks do you have?"
• The AI responds with a tool use request to call the get_drink_names tool.
• The client makes the MCP call, gets the drink names (latte, mocha, flat white), and sends them back to the AI.
• The AI then responds with "We have these drinks available: latte, mocha, and flat white."
• The user then asks "What's the price of a mocha?"
• The AI requests the get_drink_info tool with the argument "mocha".
• The client retrieves the price (6 bucks) and other information and sends it back to the AI.
• The AI responds with "Espresso and chocolate. Fantastic."

6. Key Statements

• "I would strongly recommend throughout this entire series using MCP libraries to build your servers and your clients and using the AI library to actually go and make these calls. We're just going to do that directly here because I really want you to understand how this works and to get rid of those abstraction layers so you really understand how this stuff works."
• "That's all it really takes to proxy through the MTP server tools all the way through to the anthropic AI. You just got to make that one little change in there. Otherwise the format is exactly the same."

7. Technical Terms

• MCP (Message Communication Protocol): A custom protocol for communication between the client and server.
• Tool Use: A mechanism by which an AI model requests the execution of an external function or tool.
• Input Schema: The structure and data types of the input parameters required by a tool.
• Closure: A function that has access to variables in its surrounding scope, even after the outer function has completed.
• JSON.stringify: A method to convert a JavaScript object or value to a JSON string.

8. Logical Connections

• The video builds upon the previous two parts of the series, using the DIY MCP server and client created in those videos.
• It connects the MCP client to an AI service, enabling the AI to use the MCP server's tools.
• The process involves transforming the tool definitions to match the AI's expected format and handling tool use requests from the AI.

9. Synthesis/Conclusion

The video demonstrates how to integrate a DIY MCP system with an AI service like Anthropic's Claude. It walks through the process of creating a callAI function, transforming tool definitions, handling tool use requests, and proxying MCP tool calls from the AI. By avoiding abstraction layers, the video provides a clear understanding of the underlying mechanisms involved in connecting an MCP system to an AI and enabling function calling. The coffee shop example illustrates a practical application of this integration, where the AI can use MCP tools to retrieve information about drinks and their prices.