A simple Model Context Protocol (MCP) server that provides a clean communication channel between AI agents and users.

When running automated tasks or agent workflows, it's often necessary to:

• Ask the user questions and wait for responses
• Provide status updates on long-running operations
• Communicate important information that needs user attention
• Request clarification or additional input

The speak tool provides a conversational interface for these interactions, separate from technical logs and terminal output.

Add the MCP server directly from GitHub using the Claude Code CLI:

claude mcp add github:tylerdavis/speak-mcp

This will automatically:

• Download the server from GitHub
• Install dependencies
• Build the project
• Add it to your Claude Code MCP configuration

Note: On first run, the server will automatically download and use the hfc_female (medium quality) voice as the default. You can change this at any time using the change_voice tool.

If you prefer to configure manually, add to your MCP configuration file:

{
  "mcpServers": {
    "speak": {
      "command": "npx",
      "args": [
        "-y",
        "github:tylerdavis/speak-mcp"
      ]
    }
  }
}

{
  "mcpServers": {
    "speak": {
      "command": "npx",
      "args": [
        "-y",
        "github:tylerdavis/speak-mcp"
      ]
    }
  }
}

If you want to contribute or modify the server locally:

git clone https://github.com/tylerdavis/speak-mcp.git
cd speak-mcp
npm install
npm run build

Then point your MCP configuration to the local build:

{
  "mcpServers": {
    "speak": {
      "command": "node",
      "args": ["/absolute/path/to/speak-mcp/dist/index.js"]
    }
  }
}

Sends a message directly to the user through a conversational interface using text-to-speech.

Parameters:

• message (required): The message to communicate to the user
• message_type (optional): Type of message - question, update, info, or warning

Behavior:

• Non-blocking: The tool returns immediately after generating the audio file (~500ms)
• Audio plays in the background while your agent continues working
• Perfect for real-time updates during long-running tasks

Examples:

// Ask a question
await mcp.speak({
  message: "Should I proceed with deploying to production?",
  message_type: "question"
});

// Provide a status update
await mcp.speak({
  message: "Database migration completed successfully. Starting service restart.",
  message_type: "update"
});

// Share information
await mcp.speak({
  message: "Found 3 configuration issues that need attention.",
  message_type: "info"
});

// Warn about something
await mcp.speak({
  message: "This operation will delete all existing data. Are you sure?",
  message_type: "warning"
});

Lists all available English (US) voices for text-to-speech, showing which are downloaded and which is currently selected.

Parameters: None

Example:

await mcp.list_voices();

Returns:

Available voices (16 total):

Currently selected: ryan-high (high quality, 116.2MB) *

Available to download:
1. amy-high (high quality, 116.2MB)
2. danny-high (high quality, 116.2MB)
...

Already downloaded:
• ryan-high (high quality, 116.2MB) *
• amy-medium (medium quality, 63.0MB)

Changes the active text-to-speech voice. If the voice isn't already downloaded, it will be downloaded automatically.

Parameters:

• voice (required): Voice identifier - supports multiple formats:
  • Full key: "en_US-amy-high"
  • Short name: "amy-high"
  • Partial match: "amy" (picks highest quality)
  • Numeric index: "1" (from list_voices output)

Examples:

// Change to a specific voice by name
await mcp.change_voice({ voice: "amy-high" });

// Change using partial match (picks highest quality)
await mcp.change_voice({ voice: "amy" });

// Change using numeric index from list_voices
await mcp.change_voice({ voice: "1" });

Note: First-time voice downloads take 30-120 seconds depending on voice size. Progress is displayed during download.

• Concise: Messages should be brief and to the point
• Conversational: Write as you would speak to someone
• User-focused: Avoid technical jargon unless necessary
• Context-aware: Provide enough information to be actionable

# Watch mode for development
npm run watch

# Build for production
npm run build

MIT