Fyn

I built Fyn to explore what property search could become if portals exposed their inventory to AI agents through reliable, structured tools.

Try Fyn Inspect the source

A sunlit terrace and dining table used as the cover for Fyn

I’m not trying to replace property portals with Fyn. I’m asking a narrower question: what would become possible if those portals offered official, consent-based interfaces designed for agents?

I made the current product work with public pages and portal-specific behaviour so the interaction is tangible. Its limitations show why stable, authorized agent APIs would be a better destination than teaching models to navigate interfaces built only for humans.

The hypothesisI wanted to test whether an AI could plan a property search while each portal kept control of its data, rules and final experience.

Let the model plan; keep execution deterministic

I let the AI translate a human request into locations, hard constraints and softer preferences. Fyn then validates and executes that plan without running another model in its backend.

One contract, many imperfect sources

I built a shared connector layer that maps thirteen Spanish property sources into a common listing model while preserving where every result came from.

Explain the matches—and the failures

I use explicit scoring to explain each match, and expose blocked sources, schema changes, empty results and unavailable coverage instead of hiding them.

A real MCP surface

I expose the search through a public Streamable HTTP server, so compatible AI clients can call it directly with structured errors, rate limits and an open implementation.