Since Stagehand handles the "act" part, I've been exploring whether
a "verify" layer adds value, confirming the agent is interacting
with the right element before/after action.
I built balage-core, an open-source npm library that does semantic
page analysis: given raw HTML, it returns detected endpoints (auth,
search, checkout, consent) with confidence scores and evidence chains.
MIT licensed, ~4ms heuristic mode, no browser dependency.
F1 = 66% across 20 real production websites. Alpha quality, known
gaps (Angular Material SPAs, multi-step auth).
npm: https://www.npmjs.com/package/balage-core
GitHub: https://github.com/osaka2077/balage-ainw
MCP: npx -y balage-mcp (works with Claude Desktop)
Happy to build a Stagehand integration PoC if there's interest.
Would appreciate any feedback on whether this direction is useful.
Since Stagehand handles the "act" part, I've been exploring whether
a "verify" layer adds value, confirming the agent is interacting
with the right element before/after action.
I built balage-core, an open-source npm library that does semantic
page analysis: given raw HTML, it returns detected endpoints (auth,
search, checkout, consent) with confidence scores and evidence chains.
MIT licensed, ~4ms heuristic mode, no browser dependency.
F1 = 66% across 20 real production websites. Alpha quality, known
gaps (Angular Material SPAs, multi-step auth).
npm: https://www.npmjs.com/package/balage-core
GitHub: https://github.com/osaka2077/balage-ainw
MCP:
npx -y balage-mcp(works with Claude Desktop)Happy to build a Stagehand integration PoC if there's interest.
Would appreciate any feedback on whether this direction is useful.