You do the task once in your own browser. Everything after that is editing a recording, not writing code.
01 — RECORD
Do it once, by hand.
The extension captures navigation, clicks, finalized inputs, selections, scrolls and waits, along with hints about each target element.
REC00:00:42
navigateclick ×4fill ×2selectwait
02 — EDIT
Review every step.
Rename, reorder or drop steps you didn’t mean, add waits, and pull text, lists or tables into structured output. The recording stays the source of truth.
02Select monthselect
03Open rowclick
+Extract invoice table
03 — REPLAY
Same steps, same order, every run.
The replay worker opens a fresh browser and executes the recording step by step. No model deciding what to do next.
12/12 · 14.2s
00:31Adaptive matching
The selector moved. The step still landed.
Every recorded step stores several hints about its target, not one brittle selector: role and name, labels, stable attributes, text, nearby text and position.
On replay they’re tried in order. If none finds a visible element, the run fails at that step and tells you which one.