LLM & Agents · 2026
MCP-Powered Mobility Concierge
"Is the bike actually faster than the car?" People ask in plain language. Route planners want structured input. An AI agent with the right tools bridges the gap.
→A typed MCP server + n8n agent on Gemini that compares car, bike & walking routes from plain language.
tool: compare_routes
in: { start, destination }
out: 🚗 18 min · 🚲 14 min · 🚶 41 min Highlights
- ✦Custom MCP server with one strict, typed tool
- ✦n8n agent with intent extraction & classification
- ✦Real-time results via WebSockets
- ✦Fully containerised, runs on a single box
The idea
“How do I get from Hamburg central station to the Elbphilharmonie, and is the bike faster?”
Humans ask messy questions. Route planners want clean input.
An LLM agent fits perfectly in between. If it has reliable tools.
What I built
- An MCP server (Deno + Hono). One tool: compare routes. Strict schema: start and destination, either as a place name or coordinates. It calculates car, bike and walking routes via OpenStreetMap and Google Maps and returns a comparison.
- An n8n agent on Google Gemini. Extracts intent from plain language, classifies the request, calls the MCP tool. Everything deterministic stays a plain workflow step. LLMs only where they add value.
- A real-time UI via WebSockets, with PocketBase for storage.
The lesson
Tool descriptions are prompts, too.
Being painfully explicit about input shapes made the agent dramatically more reliable.
And because it’s MCP, the same server works with Claude, ChatGPT or any other MCP client tomorrow. Build once, plug in anywhere.
