Skip to main content
Glama
README.md
# Selenium MCP Server

smithery badge

An MCP server that uses Selenium to interact with a WebDriver instance. Built using the [MCP-Server-Starter](https://github.com/TheSethRose/MCP-Server-Starter) template.

## Overview

This server allows AI agents to control a web browser session via Selenium WebDriver, enabling tasks like web scraping, automated testing, and form filling through the Model Context Protocol.

## Core Components

*   **MCP Server**: Exposes Selenium WebDriver actions as MCP tools.
*   **Selenium WebDriver**: Interacts with the browser.
*   **MCP Clients**: AI hosts (like Cursor, Claude Desktop) that can utilize the exposed tools.

## Prerequisites

*   **Node.js (v18 or later)**
*   **npm (v7 or later)**
*   **A WebDriver executable** (e.g., ChromeDriver, GeckoDriver) installed and available in your system's PATH.
*   **A compatible web browser** (e.g., Chrome, Firefox).

## Getting Started

1.  **Clone the repository:**
    ```bash
    git clone <your-repo-url> selenium-mcp-server
    cd selenium-mcp-server
    ```
2.  **Install dependencies:**
    ```bash
    npm install
    ```
3.  **Configure WebDriver:**
    *   Ensure your WebDriver (e.g., `chromedriver`) is installed and in your PATH.
    *   Modify `src/seleniumService.ts` (you'll create this file) if needed to specify browser options or WebDriver paths.
4.  **Build the server:**
    ```bash
    npm run build
    ```
5.  **Run the server:**
    ```bash
    npm start
    ```
    Alternatively, integrate it with an MCP host like Cursor or Claude Desktop (see Integration sections below).

## Tools

This server will provide tools such as:

*   `selenium_navigate`: Navigates the browser to a specific URL.
*   `selenium_findElement`: Finds an element on the page using a CSS selector.
*   `selenium_click`: Clicks an element.
*   `selenium_sendKeys`: Sends keystrokes to an element.
*   `selenium_getPageSource`: Retrieves the current page source HTML.
*   (Add more tools as needed)

## TypeScript Implementation

The server uses the `@modelcontextprotocol/sdk` and `selenium-webdriver` libraries.

```typescript
import { Server } from "@modelcontextprotocol/sdk/server/index.js";
import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js";
import { Builder, By, Key, until, WebDriver } from 'selenium-webdriver';

// Basic server setup (details in src/index.ts)
const server = new Server({
  name: "selenium-mcp-server",
  version: "0.1.0",
  capabilities: {
    tools: {}, // Enable tools capability
  }
});

// Selenium WebDriver setup (details in src/seleniumService.ts)
let driver: WebDriver;

async function initializeWebDriver() {
  driver = await new Builder().forBrowser('chrome').build(); // Or 'firefox', etc.
}

// Example tool implementation (details in src/tools/)
server.registerTool('selenium_navigate', {
  description: 'Navigates the browser to a specific URL.',
  inputSchema: { /* ... zod schema ... */ },
  outputSchema: { /* ... zod schema ... */ },
  handler: async (params) => {
    await driver.get(params.url);
    return { success: true };
  }
});

// Connect transport
async function startServer() {
  await initializeWebDriver();
  const transport = new StdioServerTransport();
  await server.connect(transport);
  console.log("Selenium MCP Server connected via stdio.");

  // Graceful shutdown
  process.on('SIGINT', async () => {
    console.log("Shutting down WebDriver...");
    if (driver) {
      await driver.quit();
    }
    process.exit(0);
  });
}

startServer();
```

## Development

*   **Build:** `npm run build`
*   **Run:** `npm start` (executes `node build/index.js`)
*   **Lint:** `npm run lint`
*   **Format:** `npm run format`

### Debugging

Use the MCP Inspector or standard Node.js debugging techniques.

## Integration with MCP Hosts

(Keep relevant sections from the original README for Cursor, Claude Desktop, Smithery, etc., updating paths and commands as necessary)

### Cursor Integration

1.  **Build your server**: `npm run build`
2.  **In Cursor:** `Settings` > `Features` > `MCP`: Add a new MCP server.
3.  **Register your server**:
    *   Select `stdio` as the transport type.
    *   Name: `Selenium Server` (or similar).
    *   Command: `node /path/to/selenium-mcp-server/build/index.js`.
4.  **Save**.

### Claude Desktop Integration

1.  **Build your server**: `npm run build`
2.  **Modify** `claude_desktop_config.json`:
    ```json
    {
      "mcpServers": {
        "selenium-mcp-server": {
          "command": "node",
          "args": [
            "/path/to/selenium-mcp-server/build/index.js"
          ]
        }
      }
    }
    ```
3.  **Restart Claude Desktop**.

## Best Practices

*   Use TypeScript and Zod for type safety and validation.
*   Keep tools modular (e.g., one file per tool in `src/tools/`).
*   Handle WebDriver errors gracefully (e.g., element not found, navigation issues).
*   Ensure proper WebDriver shutdown (e.g., `driver.quit()` on server exit).
*   Follow MCP best practices for schemas, error handling, and content types.

## Learn More

*   [Model Context Protocol Documentation](https://modelcontextprotocol.io/)
*   [Selenium WebDriver JS Documentation](https://www.selenium.dev/documentation/javascript/)
*   [MCP TypeScript SDK Documentation](https://modelcontextprotocol.io/sdk/typescript)

## Credits

**Based on the template created by Seth Rose**:

*   **Website**: <https://www.sethrose.dev>
*   **𝕏 (Twitter)**: <https://x.com/TheSethRose>
*   **🦋 (Bluesky)**: <https://bsky.app/profile/sethrose.dev>

TDQS

A3.5/5.0

Scored across 6 tools

Disambiguation5/5

Each tool has a clearly distinct purpose with no overlap: clicking, navigating, capturing snapshots, taking screenshots, typing text, and waiting for conditions. The descriptions make it easy to differentiate between them, eliminating any ambiguity in tool selection.

Naming Consistency5/5

All tools follow a consistent 'browser_' prefix with descriptive action names (e.g., click, navigate, snapshot). This uniform snake_case pattern makes the tool set predictable and easy to understand, enhancing usability for agents.

Tool Count5/5

With 6 tools, the server is well-scoped for web automation tasks. Each tool serves a specific, essential function in the Selenium domain, avoiding bloat while covering core interactions like navigation, input, and monitoring.

Completeness4/5

The tool set covers key web automation actions (navigation, interaction, monitoring), but lacks some common operations like form submission, element selection, or handling alerts. However, agents can work around these gaps using existing tools, making it mostly complete for basic workflows.

Maintenance

ActivityInactive
ResponsivenessNo issues