Mcpserver

You run an MCP server that provides real-time voice parsing, note workflows, and stock-related tooling. It exposes HTTP endpoints for programmatic control and a WebSocket stream for live voice data, wired to concurrency and scaling options to fit your environment.

How to use

You connect your MCP client to the HTTP endpoints to drive notes and parsing workflows or to the real-time voice stream for live corrections. For voice streams, you open a WebSocket connection to port 3002 and upgrade with an Authorization header carrying a JWT. Use the HTTP endpoints to authenticate, collect notes, parse them, and apply results to stock.

Key capabilities you can use include collecting notes from user input or documents, parsing those notes into structured purchases, applying parsed results to inventory, and streaming live voice chunks for on-the-fly corrections. The system supports partial parses with confidence scores and final structured outputs. You can integrate these flows with existing auth and data stores to maintain secure, auditable operations.

How to install

Prerequisites: Node.js and npm installed on your system. You should also have a functioning environment for running a small HTTP/WebSocket service. Ensure you have a JWT secret configured for secure WebSocket authentication.

1. Install dependencies for the MCP server project in your working directory.

2. Set your environment variable for JWT authentication.

3. Start the server in cluster mode for production-ready scalability.

Configuration and usage notes

The server exposes an HTTP API at the base URL http://localhost:3002 for authentication and command calls. The real-time WebSocket stream runs on port 3002 for live voice streams.

WebSocket usage: connect to ws://localhost:3002 and pass Authorization: Bearer <jwt> in the upgrade headers. Send messages as JSON with toolName live_voice_stream and appropriate parameters.

The system includes a migration mechanism for the database and supports horizontal scaling by running multiple worker processes in cluster mode. When you scale, ensure your WS connections remain sticky to each instance if you rely on session state.

Example configuration

{
  "mcpServers": {
    "http_mcp": {
      "type": "http",
      "name": "http_mcp",
      "url": "http://localhost:3002",
      "args": []
    },
    "start_cluster": {
      "type": "stdio",
      "name": "start_cluster",
      "command": "npm",
      "args": ["run", "start:cluster"]
    }
  },
  "env": [
    {"name": "JWT_SECRET", "value": "your_jwt_secret_here"}
  ]
}

Available tools

live_voice_stream

Streams live transcript chunks over WebSocket for real-time parsing and correction with partial and final results.

note_collection

Collects user or document content as a note for further parsing and processing.

parse_chicken_note

Parses a note to extract structured purchases and related data.

apply_to_stock

Applies parsed note data to update inventory stock levels.

forecast_stock

Generates stock forecasts based on historical sales data.