The integration problem
What an MCP server actually exposes
- Tools — functions the model can call (create issue, run query, send email)
- Resources — read-only context (files, schema docs, config)
- Prompts — reusable templates for common workflows
Why it matters in production
- Faster shipping — connect a Postgres or Stripe server instead of writing a new API layer per feature
- Model-agnostic design — swap Claude, GPT, or open models without rewriting tool wiring
- Clear security boundaries — hosts negotiate capabilities; sensitive actions can require human approval
How I use it today
- n8n / Make.com for scheduled workflows
- LangChain / LangGraph for multi-step reasoning
- MCP servers for live tool access during development and agent runs
