Googly
Native macOS agents that operate real apps — with your consent.
The problem
Agent demos either fake their sandboxing or silently automate the OS. Controlled native macOS automation — opening and driving apps like Keynote — safely and with visible consent is mostly unsolved in hackathon toys.
How it works
A native macOS app where you assign a specialized agent (e.g. a Pitch Deck agent) that converses, proposes actions, explains consequences, and waits for explicit consent before executing. Agents never touch macOS directly — only a thin, audited OS Bridge can, so launching Keynote fires the real “Googly wants to control Keynote” permission prompt. Agent reasoning and file generation run inside a rootless-Podman sandbox.
Key features
- Native macOS shell with chat, an Agent Monitor, and a consent UX
- Explicit OS-automation consent gates — no silent automation
- Clear sandbox boundary: agent reasoning sandboxed; OS bridge thin and audited
- Real app control — opens Keynote and prepares a deck outline
- Agent orchestrator handling selection, dialog, and consent
Architecture
Googly.app = UI + Agent Orchestrator + Sandbox Executor (rootless Podman, no OS access) + a native OS Bridge (Apple Events / NSWorkspace / ScriptingBridge). All OS actions route through the bridge only.