Things I Build

A tool-calling agent that runs on my own machine and acts on it directly. Hold a key, say what you want, and a local model decides which tools to call — driving a real Chrome session through logins, tiling every window via Win32, holding a full conversation with FinanceFlash (a flashcard app I also built) to run hands-free study sessions, reading my iCloud calendar back as a rundown, streaming the desktop to an Apple TV. A fast path pattern-matches known commands and skips the model entirely, so tiling windows or pulling today's agenda costs 0.01s and zero tokens, while anything genuinely open-ended falls through to the agent.

• Local-first, swappable brain — Ollama runs the loop on-device by default; one setting moves it to the Claude API when a task needs the extra capability. Credentials are typed in locally and never reach the model.
• Real browser, real sessions — Playwright drives an actual Chrome profile, so it inherits my logins and can carry a university SSO chain through to the far side.
• Talks to my other apps, not just the browser — a token-protected local API lets Emerald run full hands-free FinanceFlash study sessions: it asks the question, listens, grades the answer, coaches when I'm stuck, and never reveals an answer I didn't ask for.
• Writes to my real calendar, carefully — books, moves and cancels events over CalDAV, repeats included ("every Monday and Wednesday until finals"). Aim it at something that repeats and it writes nothing until it has asked whether I meant that one occurrence or the whole series — guessing wrong there quietly rewrites thirty entries.
• Learns my habits statistically — "remember this layout" logs where every window sits; once an app's position is consistent across enough snapshots, Emerald starts assuming it — no rule to write, just repeated behavior turning into a default.

Explore Emerald — full capabilities, architecture & roadmap →