自然语言优化服务器

Oxford Dictionary MCP Server

An MCP server that helps Claude write more naturally by giving it Oxford Dictionary tools to improve its own word choices while composing responses.

Quick start

bun install
bun run build

Then add the server to ~/Library/Application Support/Claude/claude_desktop_config.json and restart Claude Desktop:

{
  "mcpServers": {
    "oxford-dictionary": {
      "command": "node",
      "args": ["/absolute/path/to/dictionary_mcp/build/index.js"]
    }
  }
}

See QUICKSTART.md for a step-by-step test walkthrough.

How it works

Instead of you asking "find me a synonym", Claude automatically:

• Catches weak words like "good", "very", "thing" and improves them
• Verifies unfamiliar words sound natural before using them
• Searches for precise terms when the right word doesn't come to mind
• Matches vocabulary to the appropriate formality level

The result: Claude's writing sounds more human, varied, and natural - without you doing anything.

Tools

Four self-improvement tools:

1. find_better_word - Claude catches weak/generic words in its own writing

   • Automatically improves "good" → "effective", "excellent", "robust"
   • Prevents repetition by finding varied alternatives
   • Matches formality to conversation tone

2. get_word_examples - Claude verifies natural usage before using words

   • Checks unfamiliar words in real sentences
   • Prevents awkward or unnatural phrasing
   • Ensures words fit the context

3. check_word_meaning - Claude double-checks word meanings

   • Verifies precise definitions before use
   • Prevents word misuse
   • Ensures accurate communication

4. search_similar_words - Claude finds the exact right word

   • Searches when the precise term isn't coming to mind
   • Discovers technical or nuanced vocabulary
   • Elevates writing with sophisticated choices

Usage examples

Indirect, automatic enhancement

You type:

Write a product review for noise-canceling headphones

Claude automatically:

1. Catches "good sound quality" → improves to "exceptional audio clarity"
2. Catches "very comfortable" → improves to "remarkably comfortable"
3. Verifies "pristine" usage before writing "pristine call quality"
4. Searches for precise term → finds "passive isolation" vs "noise canceling"

You get: Natural, varied prose without generic adjectives or filler words.

More test scenarios

Creative Writing:

Write a short story opening about a mysterious house

Watch Claude enhance: "very old" → "ancient", "scary feeling" → "sense of foreboding"

Technical Content:

Explain how blockchain works in simple terms

Watch Claude verify technical terms and use precise vocabulary naturally

Professional Email:

Draft an apology to a client for a project delay

Watch Claude match formal register and avoid casual phrasing

See USAGE_EXAMPLES.md for detailed scenarios and behavior patterns.

Configuration

Claude Desktop

Add to your ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "oxford-dictionary": {
      "command": "node",
      "args": ["/absolute/path/to/dictionary_mcp/build/index.js"]
    }
  }
}

Replace /absolute/path/to/dictionary_mcp with the actual path to this directory.

Other MCP clients

The server runs on stdio transport and can be connected using:

node /path/to/dictionary_mcp/build/index.js

API credentials

This server uses the Oxford Dictionary API sandbox:

• Base URL: https://od-api-sandbox.oxforddictionaries.com/api/v2
• Authentication credentials are embedded in the code

For production use, replace with your own credentials in src/oxford-api.ts.

Design principles

Following Anthropic's best practices for agent tools:

• Token Efficient: Responses use 50-80% fewer tokens than raw API
• Natural Language: Human-readable formatting with markdown
• Clear Purpose: Each tool has a distinct, well-defined function
• Helpful Errors: Actionable error messages guide users to solutions

Development

# Watch mode for development
bun run dev

# Build for production
bun run build

The evaluation suite in eval/ runs 18 scenarios from eval/tasks.json against Claude. It needs ANTHROPIC_API_KEY and the @anthropic-ai/sdk package, which is not in package.json:

bun add -d @anthropic-ai/sdk
bun run eval/run-evaluation.ts

License

MIT

Credits

Built for TensorPoint Using Oxford Dictionary API Powered by Model Context Protocol (MCP)