My first look at Anthropic’s research preview desktop tool, including a test that helped me understand exactly where the value lies — and where it doesn’t.
Starting With a Small, Well-Defined Test
When I decided to test Claude Cowork — Anthropic’s research preview desktop tool for non-developers — I thought it might be a foundation for our video workflow system. I had recently tested Gumloop, an AI-based web app for building workflows. Gumloop didn’t succeed: I burned through all 2,000 free tokens just starting the test. That was a pretty simple test. The service runs $37 per month for 10,000 tokens which seems like it would be difficult to keep within 10k credits with a complex project. You’d also need separate accounts and tokens to run any workflow you shared, which made it impractical for team use without higher costs (and costs that would be difficult to project).
Both tools seemed to offer the same promise: automate file and task management without needing to write code or open a terminal. But as someone who uses Claude’s standard web interface regularly, I found myself asking: what does Cowork actually add?
To find out, I gave it a small, well-defined task: process receipts. I had a PDF purchasing card receipt and wanted to extract the relevant data, organize it, and produce a clean cover sheet for my TPS report — something I currently do by hand. It worked. But then it crashed, and what I learned recovering from that crash turned out to be more useful than the original test.
The Crash
At one point, the system was “thinking” and I lost connectivity for a couple of minutes. It hung. I tried quitting and restarting the app, but it would time out trying to reconnect to the API and wouldn’t recover no matter what I tried. Restarting the computer didn’t help either.
A colleague shared some diagnostic steps they’d pulled from ChatGPT — since Claude was unavailable on my Mac at that point — and I also worked through the recovery in a conversation with Claude on a different computer. I focused on how to recover my data rather than the root cause of the crash. This is a research preview, and I can understand that things will sometimes break. But the recovery conversation ended up clarifying something important about how Cowork works.
Here’s what’s happening under the hood: when you run Cowork, it spins up a lightweight Linux virtual machine directly on your Mac. This is what makes the no-terminal experience possible — Cowork has its own sandboxed environment where it can install packages, run scripts, and manage files independently. That’s genuinely impressive infrastructure running quietly in the background.
It also explains exactly what went wrong. The VM itself runs locally, but the AI reasoning that drives it happens in Anthropic’s cloud. Lose that connection mid-task, and the system can get into an inconsistent state — the local VM is still running, but the cloud side has lost its thread. For a research preview, that’s understandable. But it’s worth knowing before you depend on it for anything time-sensitive.
The Distinction That Matters Most
My initial instinct was to have Cowork build me a Python script or a Mac app that I could then run myself — or share with others. During the recovery conversation, I realized I wasn’t getting much more than I’d get from Claude.ai in a browser. Claude’s standard interface can write excellent code. You can upload files to it. You can ask it to produce scripts. The output would be nearly identical.
The meaningful difference only appears when the workflow runs inside Cowork itself. When I asked it to process my receipts, I didn’t open a terminal. I didn’t compile anything. I didn’t install dependencies. Cowork ran the Python script, encountered an error, debugged it, fixed it, and handed me finished output files — all in the background, while I waited in the chat interface. That’s the value proposition: not just writing automation, but executing it on your behalf.
This is a genuinely useful distinction for non-developers. Writing a script and running a script are two very different things if you don’t have a technical background. Cowork collapses that gap. I have years of experience loading and updating dependencies through Terminal when working with AI-generated code, and if I don’t have to do that anymore, that’s a real win.
What Cowork Adds Over the Browser Interface
There are three capabilities that genuinely differentiate Cowork from Claude.ai:
1. It executes code — no terminal required
Because Cowork runs a Linux VM locally on your machine, it has a full execution environment available at all times. When it writes a script as part of solving your task, it runs that script immediately, checks the output, and iterates. You never see the terminal. You never touch a dependency. For non-developers, the distance between “Write me something that does X” and “X is done” collapses to a single conversation.
2. It has direct access to your files
You select a folder on your computer as your workspace, and Cowork can read from and write to it directly. In the browser interface, you upload files manually, one at a time. In Cowork, I pointed it at a folder of PDFs and it worked through them. The result files appeared in my outputs folder without me touching them. For tasks involving batches of documents — receipts, reports, contracts — this changes the workflow considerably.
3. It can chain multi-step tasks without hand-holding
A task like “read these receipts, extract the data, build a spreadsheet, and generate a summary document” involves multiple steps, intermediate files, and dependencies between them. In the browser, you’d shepherd that process manually — copy output from one step, paste it into the next prompt. In Cowork, the whole chain runs in one conversation. It produces finished artifacts, not just code to run yourself.
Where It Falls Short (For Now)
As I mentioned, Cowork is explicitly a research preview, and it’s worth being clear about the real limitations I encountered — both today and likely into the near future.
It’s tied to one computer
This turned out to be more than a minor inconvenience for me. Cowork is a desktop application, and it runs its VM locally. When I lost connectivity briefly during a session, the app broke on that machine and couldn’t reconnect to the API even after restarting — it required clearing the app’s stored data entirely. More structurally, the session state and any in-progress work live on that specific machine. If you switch computers, you take your output files with you (those are just regular files), but the session itself doesn’t travel.
It’s not collaborative
Your Cowork session runs on your machine, for you. Colleagues can’t connect to it, contribute to it, or run their own tasks against it. You can share the output files — the same way you’d share any document — but the automation itself isn’t shared. For personal productivity this is fine. For team workflows, it’s a real constraint.
It isn’t cloud-synced
There’s no account-level sync of your sessions across devices. A workaround exists: if you point Cowork at a folder that’s already in iCloud or Dropbox, your output files will sync automatically. But the session context, conversation history, and workflow state don’t follow you. Each new machine is a fresh start.
An Honest Summary
Claude Cowork is genuinely useful for a specific kind of person with a specific kind of need: someone who wants to automate personal, file-based workflows on their own computer, without any technical setup. If that’s you, and you work primarily from one machine, it delivers on its promise. The receipt workflow I built runs without me opening a terminal or touching a script. That’s real.
But it’s easy to overestimate what you’re getting. If your goal is to produce a Python script, a Word document, or an analysis that you’ll use or share outside of Cowork, Claude.ai in a browser window will give you nearly the same result — and it travels with you across every device you own.
The questions I’m left with after this test are the ones I suspect the Anthropic team is also working through: What happens when the local execution model meets team workflows? Does cloud-hosted execution change the value proposition? Can session context become portable? Cowork is a research preview for a reason — these aren’t finished answers, they’re active questions.
For now, it’s a useful tool with clear edges. Understanding those edges is how you use it well.
About this post: This piece was created with the assistance of Claude and reflects firsthand testing of Claude Cowork during the research preview period (February 2026). Capabilities and limitations may change as the product evolves. Feedback can be submitted directly to Anthropic through the in-app thumbs down button.
