Fyn is a small experimental project: a demonstration of how property portals could expose inventory to AI agents through structured tools, without forcing a model to navigate an interface designed for humans.
If property inventory were available as a reliable, typed, and traceable tool, an AI could plan the search while the portal retained control over its data, rules, and final experience.
Fyn publishes search_properties through Model Context Protocol. Its schema defines municipalities, transaction type, budget, bedrooms, property types, and preferences. Free text adds context, but execution relies on validated, structured constraints.
In the prototype: MCP SDK + Zod over Streamable HTTP.The model interprets ambiguous phrases—‘near nature’, ‘neighborhood character’, or ‘renovation is fine’—and chooses locations, hard filters, and soft preferences. Fyn does not run another LLM in the backend: it receives that plan and executes it deterministically.
The separation matters: the model reasons; the tool validates and executes.A connector registry maps the common contract to each source’s actual capabilities. Today the experiment works with public pages and portal-specific behavior; ideally, these adapters would consume official APIs designed for agents.
Connectors are replaceable: the shared interface does not depend on the source.Available fields become a ListingCard: identifier, portal, URL, price, bedrooms, type, images, and observation time. Hard filters run before ranking so a high score cannot hide a basic constraint violation.
The normalized contract combines sources without erasing provenance.Explicit scoring weighs budget, bedrooms, location, and text signals such as light, views, or exterior aspect. The response adds why_matched and coverage per location and source, including blocks, schema changes, empty results, or outages.
Failure transparency is part of the result, not an internal detail.The desirable end state is official agent APIs with permissions, limits, and stable contracts.
Users keep their conversation while AI composes searches over authorized inventory.
A portal can define fields, quotas, attribution, links, access, and commercial rules.
Every property retains its source and the user completes the journey on the original listing.
Review the MCP contract and test the endpoint to see what works today and where the boundaries are.