WebDriver规范检索服务器

WebDriver Spec MCP Server

This is a Model Context Protocol (MCP) server that provides access to the WebDriver BiDi and Classic specifications. It is designed for Jetski and other MCP clients to search and read authoritative sections from the WebDriver specs.

Features

• Ingest WebDriver Specs: Fetches and parses the latest WebDriver BiDi and Classic specifications from W3C.
• Semantic Search: Search for relevant sections in the specs using natural language queries (search_specs) powered by local vector embeddings.
• Read Specs: Retrieve the full content of specific sections (read_spec_section).
• Vector Storage: Uses LanceDB for efficient local vector storage.
• Local Embeddings: Embeddings run locally on CPU via @huggingface/transformers (Xenova/all-MiniLM-L6-v2), requiring no external API keys for core operations.

Tools

The server exposes the following tools:

• search_specs: Search the WebDriver BiDi and Classic specifications for relevant sections.
  • Arguments: query (string)
• read_spec_section: Read the full content of a specific section.
  • Arguments: url (string)
• ask_webdriver: Ask a question about WebDriver specs.
  • Arguments: question (string)

Installation

1. Clone the repository.

2. Install dependencies:

   复制

   npm install
   

3. Ingest data:

   Before using the server, ingest the WebDriver specifications into the local vector store. This script fetches the specs, generates embeddings, and stores them in .mcp-data.

   复制

   npm run ingest
   

4. Build the project:

   Compile the TypeScript code:

   复制

   npm run build
   

5. Run tests:

   Verify stdio communication, tool listing, and vector querying:

   复制

   npm test
   

6. Configure MCP Client:

   For Jetski CLI, add the server configuration to your dedicated MCP configuration file at ~/.gemini/jetski/mcp_config.json (or globally at ~/.gemini/config/mcp_config.json).

   Make sure to replace <PATH_TO_REPO> with the absolute path to this repository:

   复制

   {
     "mcpServers": {
       "webdriver-spec": {
         "command": "node",
         "args": [
           "<PATH_TO_REPO>/build/src/index.js"
         ]
       }
     }
   }
   

   For other clients like Claude Desktop (claude_desktop_config.json), place the "webdriver-spec" configuration object inside "mcpServers".

Architecture

• src/index.ts: Main entrypoint setting up the MCP server over stdio.
• src/tools/webdriver.ts: Defines the MCP tools (search_specs, read_spec_section, ask_webdriver).
• src/lib/store.ts: Handles vector search and document lookup with LanceDB.
• src/lib/embedder.ts: Generates text embeddings using local transformer models.
• scripts/ingest-specs.ts: Fetches HTML from W3C, parses sections with jsdom, generates embeddings, and indexes them into LanceDB.
• test/server.test.ts: Integration test verifying clean JSON-RPC stdio communication and vector querying.

Adding a New Specification

To add a new specification that contains WebDriver BiDi or Classic endpoints, follow these steps:

1. Open scripts/ingest-specs.ts.

2. Add your specification to the SPECS array. The array contains SpecConfig objects. You need to provide:

   • url: The public URL of the specification.
   • type: A unique identifier for the spec (e.g., "new-module").
   • rootId (optional): If the WebDriver-relevant parts are a subset of the page, provide the id of the heading where the parsing should start.
   复制

   const SPECS: SpecConfig[] = [
     // ... existing specs
     { 
       url: "https://example.com/spec", 
       type: "my-new-spec",
       rootId: "automated-testing" // Only ingest sections under this heading
     },
   ];
   

3. Run the ingestion script:

   复制

   npm run ingest