Module 3 of 10, lesson 1 of 5

Tools and MCP

Follow a tool call from end to end and design tools a model uses correctly.

Much of an agent’s quality is decided by its tools.

The model executes nothing

A tool call happens in four steps:

1. You send the model the conversation and the list of tools, with their descriptions and schemas.
2. The model replies with a structured request: this tool, these arguments.
3. Your code validates that request and runs it.
4. You send the result back to the model, which continues.

Step 3 is your control point: authorization, argument validation, logging. The model proposes, your code decides.

Writing a good tool

• The model reads the name and description: they’re part of the prompt.
• Prefer a few well-separated tools to many overlapping ones.
• Return errors that explain what to fix. The model can then recover by itself.
• Limit result size: every response uses up context.
• Make writes idempotent, because an agent may repeat the same action.

MCP

Model Context Protocol is an open standard for connecting an agent to data sources and tools. A host embeds a client, the client connects to servers, and each server exposes what it can do. A tool written once works with any compatible client.

MCP solves connection, not trust. A third-party server can describe its tools misleadingly or return booby-trapped content. Treat it like a dependency: audit it.

The model proposes, your code validates and executes.

function refundOrder(input, orders) {
  const order = orders[input.orderId];
  if (!order) return { error: 'Unknown order. Ask the customer to confirm the number.' };
  if (input.amount > order.total) {
    return { error: `Amount too high. The most you can refund is ${order.total}.` };
  }
  return { refunded: input.amount };
}

const orders = { 4512: { total: 80 } };

refundOrder({ orderId: 4512, amount: 120 }, orders);

Try it in the editor

1// registry: { name: { schema, run } }2async function execute(block, registry) {3  const tool = registry[block.name];4  const output = await tool.run(block.input);5  return { type: 'tool_result', tool_use_id: block.id, content: JSON.stringify(output) };6}78module.exports = { execute };9

Legal notice and privacy Manage cookies