Fork
Fork in the middle, and vary the premise, not the attempt. One branch assumes reads dominate, one that the data can go a minute stale. Each runs far enough to show its cost, and they land side by side.
A local workbench for AI work where finding out what you want is still part of the work. Ask the model for the possibilities, not the answer. Fork them, see them side by side, challenge them, pick one. Most of it you throw away. That is how you find the one worth keeping.
Where the bottleneck went.
A tool for this has to make four moves easy.
Fork in the middle, and vary the premise, not the attempt. One branch assumes reads dominate, one that the data can go a minute stale. Each runs far enough to show its cost, and they land side by side.
Look at the work in the shape the problem has: a map of the camps, a heatmap of the contested claims, a dozen jump variants on one level. Views are yours to write, and models can already produce them.
Ask the question outside the session that is producing. A question stays a question until you decide what to do with it.
Pick from the present or from history. Throwing a version away takes it out of your current workspace, not off the shelf of options.
Runs on your machine. No cloud, no accounts. What you make is plain files in local content-addressed storage.
Generators are plain Python that can run anything: any model, any agent harness, or no AI at all. Renderers are TypeScript views for whatever they produce. Both are made in the app, and that layer is Apache-licensed and separate from it, so nothing you make is trapped.
Pre-1.0 and single-author. The hard parts are solved with care. Some of the polish isn't.
The server path needs Python 3 and uv. No account, no cloud. Don't have uv? Get it here →