Thirteen little scenes, one product: AI work on an infinite canvas that forks instead of piling up — with agents, research fan-outs and a live meter on what every branch costs.
01 · Branching
Left: each new message re-sends every earlier one, so the pile — and the bill — keeps growing. Right: a branch sends only its own short path.
A branch only re-sends its own path — up to ~90% less context re-sent (your dashboard measures the real number).
02 · Context quality
Left: as the pile grows, older turns fade out of reach and the answer goes dim — while the token bar fills. Right: two messages of clean context, a bright answer, a tenth of the tokens.
Short, clean context = better answers AND fewer tokens.
03 · Orchestration
A Project Manager agent breaks the wish into jobs, runs them at once, and assembles the result.
Describe the goal — a Project Manager agent plans, spawns workers in parallel, and assembles the result.
04 · Create app
Plan, build, launch, preview — every stage stays a node you can open and change.
From one sentence to a running app — plan, build, launch, preview, all on the canvas.
05 · Research mode
Ask once. Answers come back side by side and merge into a single cited document.
Ask once — five models research side by side and merge into one cited PDF report.
06 · Import
Bring a history export in; it comes out as a tree you can navigate and fork from.
Bring your ChatGPT / Claude history — one import turns the scroll into an organized branch tree.
07 · Model per node
Switch on one node without touching the rest of the tree.
Every node can run a different model — pick per task, not per conversation.
08 · Routing
Auto-pick sends each task where it belongs — and only pays big where big is needed.
Auto-pick routes each task to the right model — big brains for hard steps, cheap ones for easy steps.
09 · Enhance
Enhance rewrites what you typed into role, goal and constraints before it hits the model.
One click rewrites your rough ask into a precise prompt — role, goal, constraints — before it ever hits the model.
10 · Design studio
Real responsive HTML on the canvas — not a picture of a layout.
Design real, responsive HTML on the canvas — export working code, not a picture.
11 · Memory
A knowledge graph and an Obsidian vault on your own disk — recalled when it helps.
What your agents learn is kept — as a knowledge graph and an Obsidian vault on your own disk — and recalled when it helps.
12 · Providers
Cloud, local, or anything OpenAI-compatible. Keys stay on your machine.
Bring any provider — cloud keys, local models, or any OpenAI-compatible endpoint. Your keys never leave your machine.
13 · Why branching
Long single threads drift. Short branches are the documented fix.
The design follows published findings on multi-turn drift — branching is the fix, not a gimmick.
The browser playground needs no key and no signup. The desktop app is free and runs everything above.
Try in your browser ⬇ Download for desktop ▶ Watch the tutorials