Got access to Claude for Chrome after 3 months on the waitlist. First impression: it's slow as shit and asks for permission every 30 seconds, but when it works, it actually saves me hours on boring web tasks. The beta program started with just 1,000 users in August 2025 and expanded gradually to 10,000 Max subscribers.
The Reality of Getting Access
Anthropic launched this in August 2025 as a research preview for Max subscribers only. That's $200/month - yeah, seriously expensive. The Chrome Web Store listing is only accessible to approved beta users. They started with 1,000 users and expanded to 10,000, but you still need to join a waitlist and wait for approval.
The approval process took me 3 months. They're being conservative because the extension can see everything on your screen and control your browser - that's powerful and dangerous. Anthropic's safety research shows why they're cautious about releasing AI agents that can manipulate web interfaces. When you finally get approved, you get a special Chrome Web Store link via email.
Mobile support? Forget it. Desktop Chrome only - which makes sense since browser automation requires full desktop functionality that mobile browsers can't provide.
How It Actually Works
The extension sits in a side panel and takes screenshots of your active tab to "see" what you're looking at. It's basically screen scraping with AI interpretation - not as elegant as you'd hope, but it works. It basically takes screenshots and tries to figure out what to click, like a really slow human.
What it does well:
- Fills forms automatically (saved me hours on job applications)
- Clicks through multi-step workflows
- Remembers context across tabs better than most humans
- Handles dynamic content that traditional automation breaks on
What frustrates users:
- Painfully slow - 5-10 seconds per action because of screenshot analysis
- Permission overload - asks for approval every 3 seconds until you want to throw your laptop
- JavaScript confusion - React apps with dynamic loading break it every damn time. It clicks where a button used to be but isn't anymore because some useState fired and moved everything around
- Silent failures - just stops working mid-task with no error message
Claude sees your screen, figures out what elements to interact with, then sends click/type commands. Simple in theory, messy as hell in practice.
Real performance after screwing around with this thing for 2 months:
- Form filling: Works most of the time, breaks on weird sites like DMV forms
- Multi-step workflows: About half the time it gives up and I take over manually
- Data extraction: This is actually its best feature - rarely fails on tables
- Shopping/purchasing: Hell no, not testing that
The permission system is annoying as shit but necessary. You grant access per site, and it asks for confirmation before doing anything that could cost money or delete data. Makes sense when an AI can control your entire browser.
