Frontend MCP Tool Development Tutorial

MCP (Model Context Protocol) Tool Development Tutorial Table of Contents Introduction Core Concepts Usage Examples Best Practices Introduction MCP (Model Context Protocol) is a protocol for building interactions between AI models and external tools. It allows us to create custom tools and resources that enable AI models to perform specific tasks. MCP tool development uses the modelcontextprotocol SDK, GitHub repository: https://github.com/modelcontextprotocol Core Concepts 1. StdioServerTransport StdioServerTransport is the transport layer of MCP, which implements communication between server and client through standard input/output (stdin/stdout). import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio"; const transport = new StdioServerTransport(); 2. McpServer McpServer is the core server class of MCP, used to create and manage tools and resources. import { McpServer } from "@modelcontextprotocol/sdk/server/mcp"; const server = new McpServer({ name: "your-server-name", description: "Server description", version: "1.0.0" }); 3. Tool Tool is a function in MCP used to execute specific tasks. Each tool requires: Name Parameter schema (using Zod for validation) Execution function import { z } from "zod"; server.tool( "tool-name", { param1: z.string(), param2: z.number().optional() }, async (params) => { // Tool implementation return { content: [{ type: "text", text: "Execution result" }] }; } ); 4. Resource Resource is used to manage accessible data or state. Each resource requires: Name URI template Function to get content import { ResourceTemplate } from "@modelcontextprotocol/sdk/server/mcp"; server.resource( "resource-name", new ResourceTemplate("resource://{id}", { list: undefined }), async (uri, { id }) => ({ contents: [{ uri: uri.href, text: `Resource content: ${id}` }] }) ); Usage Examples Here we use the modelcontextprotocol typescript-sdk for development Install Dependencies First, install the necessary npm packages: npm install @modelcontextprotocol/sdk @playwright/test pngjs pixelmatch zod Let me demonstrate these concepts working together through a practical Playwright UI testing tool. This example will show how to create a complete visual testing tool. 1. Type Definitions First, we need to define the types for test configuration and results: // types.ts export interface VisualTestConfig { url: string; selector?: string; waitForSelector?: string; waitForTimeout?: number; threshold?: number; ignoreSelectors?: string[]; viewport?: { width: number; height: number; }; baselineImagePath?: string; baselineImage?: string | Buffer; login?: { url: string; usernameSelector: string; passwordSelector: string; submitSelector: string; username: string; password: string; successSelector?: string; }; autoLogin?: { username: string; password: string; usernameSelector: string; passwordSelector: string; submitSelector: string; successSelector?: string; loginUrlPattern?: string; }; } export interface VisualTestResult { success: boolean; message?: string; error?: string; diffPixels?: number; threshold?: number; passed?: boolean; baselineCreated?: boolean; baselineUpdated?: boolean; screenshots?: { current?: string; diff?: string; }; } 2. Create Server import { McpServer } from "@modelcontextprotocol/sdk/server/mcp"; const server = new McpServer({ name: "visual-test", description: "UI visual comparison test tool", version: "1.0.0" }); 3. Implement Core Functions 3.1 Login Function async function login(page: any, loginConfig: VisualTestConfig["login"]) { if (!loginConfig) return; await page.goto(loginConfig.url); await page.waitForSelector(loginConfig.usernameSelector); await page.waitForSelector(loginConfig.passwordSelector); await page.waitForSelector(loginConfig.submitSelector); await page.fill(loginConfig.usernameSelector, loginConfig.username); await page.fill(loginConfig.passwordSelector, loginConfig.password); await page.click(loginConfig.submitSelector); if (loginConfig.successSelector) { await page.waitForSelector(loginConfig.successSelector); } else { await page.waitForNavigation(); } } 3.2 Auto Login Function async function autoLogin(page: any, config: VisualTestConfig["autoLogin"]) { if (!config) return false; try { await page.waitForSelector(config.usernameSelector); await page.waitForSelector(config.passwordSelector); await page.waitForSelector(config.submitSelector); await page.fill(config.usernameSelector, config.username); await page.fill(config.passwordSele

Mar 27, 2025 - 07:34

MCP (Model Context Protocol) Tool Development Tutorial

Introduction
Core Concepts
Usage Examples
Best Practices

Introduction

MCP (Model Context Protocol) is a protocol for building interactions between AI models and external tools. It allows us to create custom tools and resources that enable AI models to perform specific tasks.
MCP tool development uses the modelcontextprotocol SDK, GitHub repository:
https://github.com/modelcontextprotocol

Core Concepts

1. StdioServerTransport

StdioServerTransport is the transport layer of MCP, which implements communication between server and client through standard input/output (stdin/stdout).

import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio";

const transport = new StdioServerTransport();

2. McpServer

McpServer is the core server class of MCP, used to create and manage tools and resources.

import { McpServer } from "@modelcontextprotocol/sdk/server/mcp";

const server = new McpServer({
  name: "your-server-name",
  description: "Server description",
  version: "1.0.0"
});

3. Tool

Tool is a function in MCP used to execute specific tasks. Each tool requires:

Name
Parameter schema (using Zod for validation)
Execution function

import { z } from "zod";

server.tool(
  "tool-name",
  {
    param1: z.string(),
    param2: z.number().optional()
  },
  async (params) => {
    // Tool implementation
    return {
      content: [{
        type: "text",
        text: "Execution result"
      }]
    };
  }
);

4. Resource

Resource is used to manage accessible data or state. Each resource requires:

Name
URI template
Function to get content

import { ResourceTemplate } from "@modelcontextprotocol/sdk/server/mcp";

server.resource(
  "resource-name",
  new ResourceTemplate("resource://{id}", { list: undefined }),
  async (uri, { id }) => ({
    contents: [{
      uri: uri.href,
      text: `Resource content: ${id}`
    }]
  })
);

Usage Examples

Here we use the modelcontextprotocol typescript-sdk for development

Install Dependencies

First, install the necessary npm packages:

npm install @modelcontextprotocol/sdk @playwright/test pngjs pixelmatch zod

Let me demonstrate these concepts working together through a practical Playwright UI testing tool. This example will show how to create a complete visual testing tool.

1. Type Definitions

First, we need to define the types for test configuration and results:

// types.ts
export interface VisualTestConfig {
  url: string;
  selector?: string;
  waitForSelector?: string;
  waitForTimeout?: number;
  threshold?: number;
  ignoreSelectors?: string[];
  viewport?: {
    width: number;
    height: number;
  };
  baselineImagePath?: string;
  baselineImage?: string | Buffer;
  login?: {
    url: string;
    usernameSelector: string;
    passwordSelector: string;
    submitSelector: string;
    username: string;
    password: string;
    successSelector?: string;
  };
  autoLogin?: {
    username: string;
    password: string;
    usernameSelector: string;
    passwordSelector: string;
    submitSelector: string;
    successSelector?: string;
    loginUrlPattern?: string;
  };
}

export interface VisualTestResult {
  success: boolean;
  message?: string;
  error?: string;
  diffPixels?: number;
  threshold?: number;
  passed?: boolean;
  baselineCreated?: boolean;
  baselineUpdated?: boolean;
  screenshots?: {
    current?: string;
    diff?: string;
  };
}

2. Create Server

import { McpServer } from "@modelcontextprotocol/sdk/server/mcp";

const server = new McpServer({
  name: "visual-test",
  description: "UI visual comparison test tool",
  version: "1.0.0"
});

3. Implement Core Functions

3.1 Login Function

async function login(page: any, loginConfig: VisualTestConfig["login"]) {
  if (!loginConfig) return;

  await page.goto(loginConfig.url);
  await page.waitForSelector(loginConfig.usernameSelector);
  await page.waitForSelector(loginConfig.passwordSelector);
  await page.waitForSelector(loginConfig.submitSelector);

  await page.fill(loginConfig.usernameSelector, loginConfig.username);
  await page.fill(loginConfig.passwordSelector, loginConfig.password);
  await page.click(loginConfig.submitSelector);

  if (loginConfig.successSelector) {
    await page.waitForSelector(loginConfig.successSelector);
  } else {
    await page.waitForNavigation();
  }
}

3.2 Auto Login Function

async function autoLogin(page: any, config: VisualTestConfig["autoLogin"]) {
  if (!config) return false;

  try {
    await page.waitForSelector(config.usernameSelector);
    await page.waitForSelector(config.passwordSelector);
    await page.waitForSelector(config.submitSelector);

    await page.fill(config.usernameSelector, config.username);
    await page.fill(config.passwordSelector, config.password);
    await page.click(config.submitSelector);

    if (config.successSelector) {
      await page.waitForSelector(config.successSelector);
    } else {
      await page.waitForNavigation();
    }

    return true;
  } catch (error) {
    console.error("Auto login failed:", error);
    return false;
  }
}

3.3 Visual Test Core Function

async function runVisualTest(config: VisualTestConfig): Promise {
  const browser = await chromium.launch();
  const page = await browser.newPage();

  try {
    // Set viewport size
    const defaultViewport = { width: 1280, height: 720 };
    await page.setViewportSize({
      width: config.viewport?.width ?? defaultViewport.width,
      height: config.viewport?.height ?? defaultViewport.height
    });

    // Handle login
    if (config.login) {
      await login(page, config.login);
    }

    // Visit target page
    await page.goto(config.url);
    await page.waitForLoadState("networkidle");

    // Check if auto login is needed
    await checkLoginRedirect(page, config);

    // Wait for specified element
    if (config.waitForSelector) {
      await page.waitForSelector(config.waitForSelector);
    }

    // Wait for specified time
    if (config.waitForTimeout) {
      await page.waitForTimeout(config.waitForTimeout);
    }

    // Hide elements to ignore
    if (config.ignoreSelectors?.length) {
      await page.evaluate((selectors) => {
        selectors.forEach((selector) => {
          const elements = document.querySelectorAll(selector);
          elements.forEach((el) => {
            (el as HTMLElement).style.visibility = "hidden";
          });
        });
      }, config.ignoreSelectors);
    }

    // Get page screenshot
    const screenshot = await page.screenshot({
      fullPage: !config.selector,
      type: "png",
      ...(config.selector ? { selector: config.selector } : {}),
    });

    // Save current screenshot
    const currentScreenshotPath = path.join(screenshotsDir, "current.png");
    fs.writeFileSync(currentScreenshotPath, screenshot);

    // Handle baseline image
    const baselineScreenshotPath = path.join(screenshotsDir, "baseline.png");
    if (fs.existsSync(baselineScreenshotPath)) {
      console.log("Using existing baseline image");
    } else if (config.baselineImagePath) {
      const baselineBuffer = fs.readFileSync(config.baselineImagePath);
      fs.writeFileSync(baselineScreenshotPath, baselineBuffer);
    } else if (config.baselineImage) {
      let baselineBuffer: Buffer;
      if (Buffer.isBuffer(config.baselineImage)) {
        baselineBuffer = config.baselineImage;
      } else {
        baselineBuffer = Buffer.from(config.baselineImage, "base64");
      }
      fs.writeFileSync(baselineScreenshotPath, baselineBuffer);
    } else {
      fs.copyFileSync(currentScreenshotPath, baselineScreenshotPath);
      return {
        success: true,
        message: "Created new baseline screenshot",
        baselineCreated: true,
      };
    }

    // Image comparison
    const baseline = PNG.sync.read(fs.readFileSync(baselineScreenshotPath));
    const current = PNG.sync.read(screenshot);

    if (baseline.width !== current.width || baseline.height !== current.height) {
      fs.copyFileSync(currentScreenshotPath, baselineScreenshotPath);
      return {
        success: true,
        message: "Updated baseline screenshot",
        baselineUpdated: true,
      };
    }

    // Create diff image
    const { width, height } = baseline;
    const diff = new PNG({ width, height });
    const numDiffPixels = pixelmatch(
      baseline.data,
      current.data,
      diff.data,
      width,
      height,
      { threshold: config.threshold ? config.threshold / 100 : 0.1 }
    );

    // Save diff image
    fs.writeFileSync(
      path.join(screenshotsDir, "diff.png"),
      PNG.sync.write(diff)
    );

    return {
      success: true,
      message: "Successfully created diff image",
      diffPixels: numDiffPixels,
      threshold: config.threshold || 100,
      passed: numDiffPixels < (config.threshold || 100),
    };
  } catch (error) {
    console.error("Visual comparison test failed:", error);
    return {
      success: false,
      error: error instanceof Error ? error.message : "Unknown error",
    };
  } finally {
    await browser.close();
  }
}

4. Define MCP Tool

server.tool(
  "playwright-ui-test",
  {
    url: z.string(),
    selector: z.string().optional(),
    waitForSelector: z.string().optional(),
    waitForTimeout: z.number().optional(),
    threshold: z.number().optional(),
    ignoreSelectors: z.array(z.string()).optional(),
    viewport: z.object({
      width: z.number(),
      height: z.number()
    }).optional(),
    baselineImagePath: z.string().optional(),
    baselineImage: z.string().optional()
  },
  async (params) => {
    // Get auto login configuration
    const autoLoginConfig = {
      username: process.env.AUTO_LOGIN_USERNAME,
      password: process.env.AUTO_LOGIN_PASSWORD,
      usernameSelector: process.env.AUTO_LOGIN_USERNAME_SELECTOR || "#username",
      passwordSelector: process.env.AUTO_LOGIN_PASSWORD_SELECTOR || "#password",
      submitSelector: process.env.AUTO_LOGIN_SUBMIT_SELECTOR || 'button[type="submit"]',
      successSelector: process.env.AUTO_LOGIN_SUCCESS_SELECTOR,
      loginUrlPattern: process.env.AUTO_LOGIN_URL_PATTERN || "login|signin|auth",
    };

    // Get test configuration
    const testConfig = {
      selector: process.env.TEST_SELECTOR || params.selector,
      waitForSelector: process.env.TEST_WAIT_FOR_SELECTOR || params.waitForSelector,
      waitForTimeout: process.env.TEST_WAIT_TIMEOUT ? parseInt(process.env.TEST_WAIT_TIMEOUT) : params.waitForTimeout,
      threshold: process.env.TEST_THRESHOLD ? parseInt(process.env.TEST_THRESHOLD) : params.threshold,
      ignoreSelectors: process.env.TEST_IGNORE_SELECTORS ? process.env.TEST_IGNORE_SELECTORS.split(',') : params.ignoreSelectors,
      viewport: {
        width: process.env.TEST_VIEWPORT_WIDTH ? parseInt(process.env.TEST_VIEWPORT_WIDTH) : (params.viewport?.width || 1280),
        height: process.env.TEST_VIEWPORT_HEIGHT ? parseInt(process.env.TEST_VIEWPORT_HEIGHT) : (params.viewport?.height || 720)
      }
    };

    const result = await runVisualTest({
      url: params.url,
      ...testConfig,
      baselineImagePath: params.baselineImagePath,
      baselineImage: params.baselineImage,
      autoLogin: autoLoginConfig,
    });

    if (result.success) {
      if (result.baselineCreated || result.baselineUpdated) {
        return {
          content: [{
            type: "text",
            text: result.message || "Updated baseline image"
          }]
        };
      }

      return {
        content: [{
          type: "text",
          text: `Diff pixels: ${result.diffPixels}, Threshold: ${result.threshold}, Test ${result.passed ? 'passed' : 'failed'}`
        }, {
          type: "image",
          data: result.screenshots?.current || "",
          mimeType: "image/png"
        }, {
          type: "image",
          data: result.screenshots?.diff || "",
          mimeType: "image/png"
        }]
      };
    }

    return {
      content: [{
        type: "text",
        text: result.error || "Unknown error"
      }]
    };
  }
);

Project Repository

https://www.npmjs.com/package/@anguske/mcp-playwright-visual-test

5. Configure Cursor

To use this tool in Cursor, add the following configuration to .cursor/mcp.json:

{
  "mcpServers": {
    "@anguske/mcp-playwright-visual-test": {
      "command": "npx",
      "args": ["-y", "@anguske/mcp-playwright-visual-test"],
      "env": {
        "AUTO_LOGIN_USERNAME": "",
        "AUTO_LOGIN_PASSWORD": "",
        "AUTO_LOGIN_USERNAME_SELECTOR": "#userNameSignIn",
        "AUTO_LOGIN_PASSWORD_SELECTOR": "#passwordSignIn",
        "AUTO_LOGIN_SUBMIT_SELECTOR": "input[type=\"submit\"]",
        "AUTO_LOGIN_SUCCESS_SELECTOR": "",
        "TEST_VIEWPORT_WIDTH": 1440,
        "TEST_VIEWPORT_HEIGHT": 800,
        "TEST_THRESHOLD": 20,
        "TEST_WAIT_TIMEOUT": 10000,
        "AUTO_LOGIN_URL_PATTERN": "login|signin|auth",
        "PROJECT_ROOT": "C:/project/root"
      }
    }
  }
}

This configuration file defines the MCP server configuration information:

Server Configuration:
- command: Use npx command to run the tool
- args: Use -y parameter to automatically confirm installation and run the @anguske/mcp-playwright-visual-test package
Environment Variable Configuration:
- Auto login configuration:
  - AUTO_LOGIN_USERNAME: Login username
  - AUTO_LOGIN_PASSWORD: Login password
  - AUTO_LOGIN_USERNAME_SELECTOR: Username input field selector
  - AUTO_LOGIN_PASSWORD_SELECTOR: Password input field selector
  - AUTO_LOGIN_SUBMIT_SELECTOR: Submit button selector
  - AUTO_LOGIN_SUCCESS_SELECTOR: Login success indicator selector

Test configuration:
- TEST_VIEWPORT_WIDTH: Viewport width (1440px)
- TEST_VIEWPORT_HEIGHT: Viewport height (800px)
- TEST_THRESHOLD: Difference threshold (20)
- TEST_WAIT_TIMEOUT: Wait timeout (10000ms)
Other configuration:
- AUTO_LOGIN_URL_PATTERN: Login page URL matching pattern
- PROJECT_ROOT: Project root directory path

When you use this tool in Cursor:

Cursor reads this configuration file
Starts the MCP server according to the configuration
Runs tests using the configured environment variables
Displays test results and screenshots

For example, when you enter in Cursor:

/test playwright-ui-test url="https://example.com"

Cursor will:

Recognize this as a tool call
Start the server using configured environment variables
Call the playwright-ui-test tool method
Display test results and screenshots

6. Add Resource

server.resource(
  "testResult",
  new ResourceTemplate("test://{id}", { list: undefined }),
  async (uri, { id }) => ({
    contents: [{
      uri: uri.href,
      text: `Test Result ID: ${id}`
    }]
  })
);

Best Practices

Parameter Validation
- Use Zod for strict parameter validation
- Provide default values for optional parameters
- Use environment variables for configuration
Error Handling
- Use try-catch to catch possible errors
- Return structured error information
- Clean up resources in finally block
Resource Management
- Use meaningful URI templates
- Implement appropriate resource access control
- Manage temporary files and directories
Tool Design
- Keep tool functionality single-purpose
- Provide clear parameter documentation
- Return structured results
- Support multiple configuration methods
Configuration Management
- Use environment variables for configuration
- Provide reasonable default values
- Support multiple configuration sources

Summary

MCP provides a powerful framework for building interactions between AI models and external tools. Through proper use of Tools and Resources, we can create feature-rich and maintainable AI applications. Remember to follow best practices to ensure code maintainability and extensibility. In actual development, pay attention to:

Proper code organization
Comprehensive error handling
Flexible configuration options
Code testability
User experience and feedback