Module 5: Model Context Protocol (MCP)

Lesson 5: Tool Discovery and Invocation

In the final lesson of Module 5, we look at the Automation of tool use. In a custom agent (Module 4), you spend a lot of time writing code to map "Model JSON" to "Python Function." MCP makes this logic Invisible.

In this lesson, we deconstruct the "Discovery-to-Invocation" pipeline that makes Claude-Desktop and other MCP clients so powerful.

---

1. Automated Discovery

When you start an MCP Client, it iterates through all configured servers.

1. It calls tools/list on every server.
2. It aggregates these into a single "Universal Tool Set."
3. It sends this set to the model's System Prompt.

Why this is a "Game Changer":

You can add a new tool to your server and Restart the Client. The model immediately uses it. There is no "Update Prompt" or "Modify Code" step in the client application itself.

---

2. Standardized Invocation

When Clause decides to call a tool, it uses the standard JSON-RPC Request format.

{
  "method": "tools/call",
  "params": {
    "name": "read_file",
    "arguments": {
      "path": "src/main.py"
    }
  }
}

The Client's Role:

The Client doesn't need to know read_file. It just looks at its "Routing Table" (built during Discovery) to see which server owns read_file and forwards the message.

---

3. The Tool Result Life-Cycle

The result from an MCP tool is more than just a string. It is an Object containing:

• content: The actual data (Text or Image).
• isError: A boolean flag telling the client if something went wrong.

Architect's Rule: Always return isError: true for technical failures (e.g., File not found). Do not just return a string saying "Error." If the Client sees isError: true, it can trigger built-in Retry Logic (Module 10, Lesson 2).

---

4. Visualizing the Pipeline

graph LR
    A[Add Server] --> B[Auto-Discovery]
    B --> C[Model Awareness]
    C --> D[Model Tool-Intent]
    D --> E[Standard Invocation]
    E --> F[Standard Result]
    F --> G[Model Reflection]

---

5. Summary of Module 5

Module 5 has centered on Standardization.

• You learned why MCP is the protocol of choice (Module 5, Lesson 1).
• You mastered the Discovery Handshake and Message types.
• You explored Distributed Architectures (Stdio vs. HTTP).
• You used Resources for Context Injection.
• You deconstructed the Automation of Discovery/Invocation.

In Module 6, we look at how to use these protocols in the world's most advanced AI coding environment: Claude Code Configuration and Workflows.

---

Interactive Quiz

1. How does MCP eliminate the "Glue Code" between the model and the tool?
2. What is the "Routing Table" in an MCP client?
3. Why should an MCP tool return an isError boolean instead of just an error message string?
4. Scenario: You have 10 MCP servers. How does the model know which server to call for a specific task?

---

Reference Video: