A real IDE underneath.
Not a wrapper around a chat box — terminals, git, files, and builds are first-class.
Terminals
Real ptys, not simulations
node-pty terminals with WebGL rendering: the shell draws its own prompt, colors and interactive programs just work. When an agent runs a command, its terminal appears beside it — watch, or take over.
❯ claude # boots signed-in, in your workspace
❯ git worktree add ../fix-123 # isolated agent sandboxes
Agent registry
Ten CLIs, one + menu
Built-in presets for the agents you already pay for — each runs the official CLI with your existing install and auth. Add anything else as a custom entry, ACP or terminal.
Claude CodeCodexGeminiOpenCodeQwenAiderGooseAmpKimiCrush
Git
Commit beside the agent that made the change
A commit panel card: browse changes, stage hunks, side-by-side diffs, commit — without leaving the window. Checkpoints snapshot the working tree before every agent turn, so any turn is one click to rewind.
Safety
Permission cards, not YOLO mode
Agent asks are inline cards with allow-rules you control; sensitive-file globs are enforced below the agent — it can't read what you've fenced off. Autonomy is a ladder (Observe → Propose → Execute), not a switch.
Writing
LaTeX mode
Open a .tex workspace and the shell leans into paper-writing: latexmk builds, PDF preview with SyncTeX jumps, and Overleaf sync.
Keyboard-first
The whole shell on home row
⌘L talk to any agent · ⌘K palette · ⌘P files · ⌘J cards · ⌘1–9 sessions · ⌃Tab cycle — every chord rebindable from keymap.json.
