LifeMD AI Proxy

Express proxy that hosts the LifeMD multi-agent workflow. It exposes HTTP endpoints that forward user text or voice prompts to an OpenAI-powered agent that can route health-related questions to Model Context Protocol (MCP) tools for women's health, weight management, and mental health advice.

Features

• @openai/agents LifeMD agent wired up with three MCP tools (womens_health_tool, weight_management_tool, mental_health_tool).
• REST endpoint (POST /api/ai) for direct text prompts.
• Voice endpoint (POST /api/voice) that accepts an audio file, transcribes it with OpenAI (gpt-4o-mini-transcribe), and funnels the transcript to the agent.
• TypeScript source with build step that emits ESM JavaScript to dist/.

Prerequisites

• Node.js 20+
• npm 10+
• An OpenAI API key with access to GPT-4o Mini (chat + transcription models)

Environment Variables

Create a .env file or export the variables before launching the server.

Variable	Description
OPENAI_API_KEY	Required. Used by both the HTTP proxy and the MCP tools/agent.
PORT	Optional. Defaults to 8080.

Install & Run

npm install           # install dependencies
npm run dev           # start TS server with ts-node (hot reload)
npm run build         # compile TypeScript to dist/
npm start             # run compiled server (build runs automatically first)

API

POST /api/ai

• Body: { "message": "string" }
• Response: { "answer": "string" }
• Errors: 400 when message missing/empty, 500 for unexpected agent issues.

POST /api/voice

• Content-Type: multipart/form-data
• Field: audio (single file, e.g. .webm, .wav)
• Response: { "transcript": "text", "answer": "string", "audio": { "mimeType": "audio/mpeg", "base64": "..." } }
  • audio.base64 can be converted into a data URL on the frontend (data:${mimeType};base64,...) for immediate playback.
• Errors: 400 when file missing, 500 if transcription or agent call fails.

Development Notes

• Source entry point: src/index.ts; build target: dist/index.js.
• MCP server entry: src/mcp/index.ts (compiled to dist/mcp/index.js). The agent dynamically picks the compiled version when present, otherwise falls back to ts-node.
• The /api/voice endpoint depends on multer for in-memory uploads and openai/audio.transcriptions.create with toFile helper from the OpenAI SDK.

Code Structure Cheat Sheet

• src/index.ts – bootstraps Express, mounts shared middleware, and wires the modular routers.
• src/routes/ – houses one router per HTTP surface (ai.ts for text prompts, voice.ts for uploads). Each router handles its own middleware stack (e.g., Multer for voice) and uses shared helpers.
• src/utils/ – reusable HTTP helpers (http.ts) and validation logic (validation.ts) that enforce consistent error messaging and request limits.
• src/config.ts / src/errors.ts – central place for environment-driven settings and HTTP-friendly error types.
• src/mcp/ – contains the MCP server exposed to the agent. openai.ts wires the shared SDK client, tools/ holds one file per MCP tool plus a createAdviceTool factory, and index.ts simply registers everything.
• src/agent/ – mirrors the same idea for the agent runtime. openai.ts validates the API key and sets the default key for @openai/agents, mcpServer.ts handles child-process lifecycle + shutdown hooks, lifeMdAgent.ts defines the agent persona, and index.ts exports runLifeMdAgent for the HTTP layer.
• When adding a new tool or agent behavior, drop a new file in the respective folder and export it through the local index.ts barrel to keep each concern isolated.

Testing Locally

1. Ensure OPENAI_API_KEY is set (export OPENAI_API_KEY=...).
2. Run npm run dev and open another terminal.
3. Test text flow:

   curl -X POST http://localhost:8080/api/ai \
        -H 'Content-Type: application/json' \
        -d '{"message":"I have a headache"}'

4. Test voice flow (replace clip.webm with your file):

   curl -X POST http://localhost:8080/api/voice \
        -F '[email protected]'