LeadFunnel MCP — Render

MCP server with the session in the process RAM. No database, no disk, no cache.

POST /cargar   →  Salesforce → depura → scorea → RAM
POST /mcp      →  Claude consulta (lee de RAM)
POST /purgar   →  RAM = null
GET  /health   →  ¿está vivo? ¿hay sesión?

When Render shuts down the service due to inactivity, the process dies and the data goes down with it. That's the ephemerality guarantee — it doesn't depend on a TTL.

Deploy

1. Push this repo to GitHub (private).

2. Render → New → Web Service → connect the repo.

3. Settings:

   • Runtime: Node

   • Build command: npm install

   • Start command: npm start

   • Health check path: /health

Environment variables

Related MCP server: Salesforce MCP Server

Testing

# Salud
curl https://TU-APP.onrender.com/health

# Cargar (tarda: hace todo el pull de Salesforce)
curl -X POST https://TU-APP.onrender.com/cargar \
  -H "Authorization: Bearer $MCP_TOKEN"

# Handshake MCP
curl -X POST "https://TU-APP.onrender.com/mcp?k=$MCP_TOKEN" \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"curl","version":"1"}}}'

# Purgar
curl -X POST https://TU-APP.onrender.com/purgar \
  -H "Authorization: Bearer $MCP_TOKEN"

The Accept header with text/event-stream is mandatory in Streamable HTTP. Without it, the SDK rejects the request.

Official inspector before connecting to Claude:

npx @modelcontextprotocol/inspector

Connect to Claude

claude.ai → Customize → Connectors → + → Add custom connector

URL: https://TU-APP.onrender.com/mcp?k=TU_TOKEN

Daily workflow

1. POST /cargar when you start

2. Query whatever you want from Claude

3. If you've been away for a while, Render shut down the service: hit /cargar again

Known limitations

Cold start. On the free tier, after ~15 min without traffic Render shuts down the service. The next request wakes it up (~50 s) and the session is gone. If that's a problem: paid plan, or a cron job hitting /health every 10 min (although that keeps data alive in RAM indefinitely, which is exactly what you didn't want).

Token in the URL. The Claude connector UI doesn't allow arbitrary headers. The token in a query param is weaker than a header: it stays in the connector config and potentially in logs. The next step is real OAuth 2.0, which is doable with the SDK but is another project.

RAM. 12k leads ≈ 25 MB in JS objects. Render's free tier gives you 512 MB. Plenty of headroom, but if the dataset grew a lot you'd need to paginate.

What is stored

With ENMASCARAR_PII=false, full phone numbers and emails are stored. It's a conscious risk decision: with that variable set to false, real candidates' contact data enters the model context on every query.