I’ve open-sourced an MCP server that lets an AI assistant answer presence questions in plain language: GitHub - positionguard/positionguard-mcp: MCP server for PositionGuard: area-level presence for AI assistants, read-only, no coordinates. Answers only what a member has chosen to share, and says 'unknown' otherwise. · GitHub (MIT, Node/TypeScript)
Who is home?
No one is confirmed at Home right now — but note that 5 members of the “My Family” group haven’t shared their presence with assistants or don’t have a fresh location report, so any of them could actually be there. I can’t confirm either way for those members.
That’s a real session against Open WebUI. It also runs under Claude Desktop over stdio. Three transports, so it should drop into whatever you’re already using.
Three things in there worth pulling out.
It answers in areas, not coordinates. Home, the Gym, the named areas you’ve already set up. The assistant never receives a latitude and longitude, because the key it holds can’t ask for them. That’s a scope on the key, not a promise in the docs.
The people don’t have to live with you. “Runners Club @ Gym” is people from my running club, not my household, and the Gym isn’t my house. Presence detection normally stops at your own walls and your own devices. “Who’s at the office?” works the same way.
It tells you what it doesn’t know. Sharing with an assistant is a separate opt-in from sharing with the group, so an answer is always bounded by what people actually agreed to. I’d rather it say that than guess.
There’s also nothing to configure per person, no YAML, no per-member entries. Someone joins a group and they’re simply there the next time you ask.
The part I didn’t expect to enjoy: I never wrote a rule for any of those questions.
Happy to answer questions on setup, or on how the scoping works if that’s the interesting part for you.