Transform
file in → ask in plain English → result file out. A small model only picks the operation; the same tools then run either on the SpeedyDb micro server, which streams so a file larger than memory is answered slowly rather than not at all, or entirely in this browser, where the file is never uploaded at all. model · readyYour files live in a workspace of your own on the server and are visible to nobody else. “New workspace” starts an empty one — the old files are then only reachable by the key that made them, and are swept after a period of inactivity.
Pick one file to transform it. Pick two or more and “Combine selected” stacks them end to end into one file, which then becomes the file the next steps run against. Combining is a button rather than something to ask for, because selecting the files has already said it.
Readable here: .csv, and the .spdb datasets this server builds
from them. Other formats are refused rather than loaded — a Parquet file larger than
memory would defeat the point of a server whose whole promise is that it never loads one.
Chain steps with “then”. Every plan is labelled with who made it — tools is a native tool call, model is JSON recovered from prose, and rules means the model failed and keywords took over.
ready.