Google Just Announced My Side Project

I was mid-implementation on a meeting agent when the Google Meet support page dropped — new AI features, real-time assistance during calls, the whole thing. The notification arrived while I had the code open.

This happens. You're not special when it happens to you.

The thing is, Google's approach is the hard way. Cloud transcription, their infrastructure, their models, their latency, their data. The Cluely approach — a desktop app that monitors system audio directly — is cleaner in every dimension that matters. You get the audio before it leaves the machine. You process it with local Whisper in real time. Nobody's servers touch it. The latency collapses.

From there the architecture is obvious. Local SQLite, chunked by time, updated continuously as Whisper flushes segments. The agent gets a live transcript database — not a stream, not an API, a queryable database — and can answer questions about what was just said, what was said twenty minutes ago, or what was said in a meeting three months ago, because the semantic index spans everything and lives centrally.

That last part is the part Google can't do. They get your meeting. You get your meetings, plural, as a single corpus that your agent can reason across. Someone references a decision from Q3 — your agent was there.

Sheldon, which I've been building as an iOS app, is already almost this for macOS. The bones are right. Desktop app, system-level audio access, local inference — the platform just changes.

Linux is fine, honestly. Linux people will make it work and be quietly smug about it.

Windows, I'm sorry. That's a 2060 problem.