geppetto is a native macOS app for running AI coding agents — Claude Code, Devin — with a real interface instead of a terminal window.

The agents themselves work fine. The problem is that a terminal is a bad place to review what they did.

Making agent work legible

An agent working in a terminal produces a scrolling wall of ANSI text: what it read, what it changed, what it is about to run, all flattened into one stream at one level of importance. You can't skim it, you can't collapse it, and the one line that actually needed your attention went past four seconds ago.

geppetto renders the same session as native UI. Tool calls become collapsible cards. Diffs render as diffs. Thinking folds away. Permission requests arrive as sheets you approve deliberately, rather than as another prompt you've already dismissed by muscle memory.

Parallel by default

Each session gets its own tab and its own working directory. Agents sharing a project get isolated git worktrees, so two of them can work the same repo without fighting over the index.

The interface isn't a wrapper around the terminal. It's the argument that agent output deserves to be designed.