implementing-mcp-prompts

Implementing MCP Prompts

Workflow

Task Progress:
- [ ] Step 1: Create prompt file in src/mcp/prompts/
- [ ] Step 2: Define Zod schema, name, config, and callback
- [ ] Step 3: Export as default PromptPlugin
- [ ] Step 4: Register in src/mcp/prompts/mod.ts
- [ ] Step 5: Run `deno task ci` to verify

Prompt File Template

Every prompt follows this structure. Create a new file in src/mcp/prompts/.

import type { GetPromptResult } from "@modelcontextprotocol/sdk/types.js";
import { z } from "zod/v3";

import type { PromptPlugin } from "$/shared/types.ts";

// 1. Validation schema
const schema = z.object({
  myArg: z.string().min(1, "Required"),
});

// 2. Prompt name (kebab-case)
const name = "my-prompt-name";

// 3. Config — choose one of the two argument styles below
const config = {
  title: "Human-readable title",
  description: "What this prompt does — shown to the LLM",
  argsSchema: {
    myArg: schema.shape.myArg,
  },
};

// 4. Callback — returns GetPromptResult with messages array
async function callback(args: Record<string, unknown>): Promise<GetPromptResult> {
  const { success, data, error } = schema.safeParse(args);
  if (!success) throw new Error(error.message);

  return {
    messages: [{
      role: "user",
      content: {
        type: "text",
        text: `Do something with: ${data.myArg}`,
      },
    }],
  };
}

// 5. Export as PromptPlugin tuple
const prompt: PromptPlugin = [name, config, callback];

export default prompt;

Registration

In src/mcp/prompts/mod.ts:

1. Import the prompt: import myPrompt from "./myPrompt.ts";
2. Add it to the prompts array:

export const prompts: PromptPlugin[] = [
  // ... existing prompts
  myPrompt,
];

The server registers each prompt via server.registerPrompt(...prompt).

Key Type

From src/shared/types.ts:

export type PromptPlugin = Parameters<McpServer["registerPrompt"]>;

This is a tuple: [name, config, callback] matching McpServer.registerPrompt() parameters.