Thanks to Claude Cowork, I’m finally at Inbox Zero

TL;DR: Anthropic’s new AI agent, Claude Cowork, has impressed industry experts, journalists, and productivity obsessives with its ability to carry out multistep computer tasks on your behalf. But for me, it’s not so much a coworker as it is a professional cleaner you’ve hired to organize your messy digital home.

What happened: Claude Cowork, released earlier this week, is Anthropic’s buzzy new agentic AI tool that bridges the gap between the developer-focused Claude Code and what a typical office worker might need. (Anthropic also boasts that Claude Code wrote Cowork in under two weeks, with a little help from human software engineers.)

The bottom line: Cowork marks a shift from chat-only AI assistants to one that lives and operates on your computer—as long as you give it access to everything. Anthropic just made it available to all Pro subscribers, so you can get it for $20 a month (we had to fork out $100). For your money, Cowork can read, edit, create, and organize your files. (It can also delete them if you’re not careful.) It didn’t overhaul how I use my computer day to day, but it’s a glimpse of how people might use AI in their daily workflows in the very near future.

Where Claude Cowork was great:

• Organizing my chaotic desktop by file type—dumping screenshots, PDFs, and Word documents into separate folders—without touching existing folders.
• Deleting and organizing emails. We live in email-slop hell; I had over 10,000 messages in Gmail, including more than 4,000 unread. I gave Claude very detailed rules for what to keep and what to delete. It took an hour or two, but my Gmail storage usage dropped from 64% to 31%.
• Intelligently renaming all my screenshots based on what they depict.
• Searching my emails for receipts from the past month (both in email bodies and attachments), downloading them into an Expenses folder, and tallying the spending in a spreadsheet. It wasn’t fast, but it did it accurately.
• Finding the best movie showing. I asked Claude to find the most central theater for my friend group, with seats in the middle rows. It found a time and place that met all my parameters.

Where it was so-so:

• Identifying nuanced email types. When I asked Claude to label PR pitches, it initially lumped newsletters in too. After clarification, it improved but still didn’t catch all of them.
• Handling multiple Google accounts. Claude sometimes got confused about which account it was pulling from, even with the correct one open in a tab.
• Compressing large video files. Compression is slow in general, but Claude was glacial—it didn’t even finish one after about an hour. About 30 minutes in, it noticed the snail’s pace and switched compression methods. It also couldn’t replace the originals due to a permission issue, so I had to clean up manually.
• Granting folder access. Sometimes Claude gave up if it lacked access; other times it asked me to point it to the right folder, or found it and prompted me to approve access.

Where Claude fell flat on its face:

• It could not find a PDF I downloaded earlier that day and summarize it in five bullet points. The only response: “PDF too large. Try reading the file a different way (e.g., extract text with a CLI tool).”
• It wasn’t able to filter top news stories. I asked Claude to identify major tech news from the past 24 hours. Aside from its questionable news discernment, it kept surfacing stories from days earlier.
• It failed to do simple macOS tasks without extra setup. For example, adding a to-do list to the Reminders app required installing Claude Code (a whole process) and issuing the command there (easy once installed). At first, Claude merely gave me instructions on how to open the Reminders app or give a voice command to Siri like I was its boomer parent. Thanks, Claude!

Conclusion: My time with Claude Cowork was a nice collaboration, but not worth $100 a month (for $20, however, it could be worth it). Using it felt like a preview of what could come next—something closer to an AI agent baked into a device, without the need for endless connectors and extensions. (Lenovo and Apple are both attempting to add agentic AI to devices.) With Cowork, nearly every step triggers a permission pop-up, which is understandable for security reasons but cumbersome nonetheless. True seamless integration isn’t here yet—and unlike a real coworker, you can’t exchange an annoyed look with Claude when the all-hands runs too long. —WK