BrowserClaw – Accessibility snapshot and ref targeting for AI browser agents

github.com

2 points·MrRubin··4 comments

Most AI browser tools make the LLM do the hard part: interpret a screenshot and guess coordinates, or generate CSS selectors that break on the next deploy.

BrowserClaw flips it. Snapshot the page's accessibility tree, get a numbered text map, AI picks a ref, Playwright resolves it to the exact element. The AI does what it's good at (reading text), the library does what it's good at (hitting elements reliably). No vision model, no selectors, no guessing.

Extracted from OpenClaw (github.com/openclaw/openclaw) and packaged as a standalone library -- battle-tested CDP/Playwright internals plus ongoing community fixes, without pulling in the full runtime. Drop it into any agent loop.

npm install browserclaw https://github.com/idan-rubin/browserclaw

For anyone interested in this becoming an official shared foundation with OpenClaw, there's an open discussion here: https://github.com/openclaw/openclaw/discussions/15804

this is fire. by the way, the snapshot extraction method of browser-use looks interesting- any thoughts on that? not only it assigns a ref to the tree, it also captures the difference between the previous action

Thanks! browser-use's diff approach is clever for reducing noise, but in practice full snapshots are more reliable for long multi-step workflows — no state drift, no edge cases when a diff misses a dynamic update. Simpler mental model for the LLM too.

Reply on news.ycombinator.com