I run ads for a small travel business. I started using Claude Code with Claude in Chrome to do real work inside our logged-in Meta Ads Manager: read results by date, split location targeting between two ad sets, publish, and collect the ad post links for our team.
The usual agent loop is screenshot → guess → click → screenshot. On Ads Manager's heavy editor, the screenshot call timed out at 30 seconds six times in one session. What kept working was reading the page as text and the accessibility tree — the same structure screen readers and Windows UI Automation use. The agent found the "Remove" button for one place by name, clicked it by reference, read the list back as text, and published.
So I packaged that habit into ChromeFast.
Not a new idea — a new focus
Reading the accessibility tree isn't new: Microsoft's Playwright MCP is built on accessibility snapshots. ChromeFast builds on that idea for a specific job: your real logged-in browser and business dashboards, with one rule set that works in Claude Code, Codex, Cursor and Antigravity, plus safety gates before anything that spends money or publishes.
The numbers (reproducible)
I benchmarked 5 public pages in real Chrome:
- Reading the whole page as text cost 2.6×–37× fewer tokens than the screenshots needed to see it, and was 5.6×–27× faster to capture than a single screenshot.
- Locating one button by role/name cost ~9–470 tokens, versus ~1,366 for every screenshot.
- Honest surprise: dumping the full tree can cost more than a screenshot (10k–70k tokens on big pages). So the rule is text first, then a scoped find — never the whole tree.
What's in the repo
- Claude Code skill + plugin,
page-readerandpage-operatoragents,/chromefast:readand/chromefast:do - Dashboard recipes: Meta Ads Manager (live-tested), Google Ads, GA4, Search Console, Shopify admin (testers wanted)
-
AGENTS.mdfor Codex/Antigravity/Windsurf, Cursor rules, installers - the benchmark, so you can check the numbers yourself
👉 https://github.com/umarahad2005/chromefast — if it saves you time, a ⭐ helps others find it. And if you can test a 🧪 recipe on your own dashboard, I'd love a PR.
I wrote this with help from Claude (AI); the benchmark numbers come from a script anyone can rerun.
Top comments (0)