I’ve been exploring modes of MCP development of a Screentime MCP server. In the loops, I ask it to look at my app and browsing behavior and summarize it. Obviously very sensitive, private information.

Claude generated SQL and navigated the data and created a narrative of my poor behavior patterns including anxiety-ridden all-night hacking sessions.

I then asked it if it considered time zones … “oh you’re absolutely right! I assumed UTC” And it would spit another convincing narrative.

“Could my app-switching anxiety you see be me building in vscode and testing in ghostty?” Etc

In this case I’m controlling the prompt and tool description and willfully challenging the LLM. I shudder to think that a desperate person gets bad advice from these sorts of context failures.

▲

iJohnDoe 3 days ago | parent [-]

Which Screentime app/source?

	▲	neomantra 2 days ago \| parent [-]
		You can explore with it here [1]. Still an experiment-in-progress, but it is OSS so anyone can play along. Some agentic loop scaffolding is in there. Another cool thing is that I was in two simultaneous dev loops, one working on this and the other working on Charm Crush to work on things like this. I’m trying to figure out the tool to make these “bridge structured data and LLM” MCP tools. I’m circling back to the AgentDank cannabis data work with these learned lessons (same GH org) and then onto the abstraction of all this. [1] https://github.com/AgentDank/screentime-mcp