Web Draw: drive a real browser from a text-only model, Amazon page in ~750 tokens

ahstanin · reddit · 2026-08-28

The author released Web Draw, a free Chrome extension that removes the vision requirement from browser automation. Most approaches assume screenshots, which rules out text-only models and costs several thousand tokens per observation for those that can see. Web Draw renders the visible page as text with a stable handle on every control — observe, act by handle, observe again — a loop a 7B/8B text model can run.

Pages render as control listings like e18 textbox "Tracking Number" required invalid=...: an Amazon search page is 750 tokens, a full checkout 550. Data tables render as markdown, repeated structures like feed posts collapse into groups, and an off-screen line tells the model what's above/below the fold.

Since small models fail differently from large ones, most of the work went into removing ambiguity: a control covered by an overlay is flagged rather than clicked; an ambiguous target fails with matching candidates listed rather than silently picking one; a refused submit reports what the page said instead of looking like success. It runs in your normal browser with existing logins, talks only to 127.0.0.1, free, no account.

Related event: Web Draw Renders Pages as Text So Text-Only Models Can Browse(2 posts)→

Original post →

More from coding & agent

coding & agent channel →