Fredy, my self-hosted flat finder, can now be set up by talking to a localLLM instead of filling in the form

Fredy watches 24 real estate portals across Germany, Austria, Switzerland, Italy, Spain and Portugal, throws away duplicates, and notifies you via Telegram, ntfy, Discord, Slack or email. No cloud, no account, no API key. It has had an MCP server for a while, so you could point your agent of choice at it and ask what turned up this week. Read-only until now. New: it can write back. Your assistant can add notes to a listing, put one on the watchlist, take it off again, and set up a whole new search. The last one took the longest. A search needs a name, the URLs you copied out of the portals, rent or buy, and a channel to notify. I first handed all of that to the model in one tool call and let it collect the answers. That works with big LLM's... and is useless with a 7B model, which guesses instead of asking. So Fredy runs the interview itself. One question at a time, in your language, each answer checked as it comes in. Paste a portal homepage instead of a search result and it tells you. Name a channel you do not have and it lists the ones you do. Nothing gets created until it has read the summary back and you said yes. It will not draw an area filter and it will not create a notification channel. The first needs a map, the second would mean sending a bot token through the model. Both stay in the web UI. Repo: https://github.com/orangecoding/fredy Demo: https://fredy-demo.orange-coding.net