I’ve been running Claude.AI Desktop as a lightweight “AI OS” on my Mac for months. The connector ecosystem turned it into my command center—Linear integration for project management, Google Drive for document access, the whole MCP stack piping context wherever I need it. Works well. Mostly.
The crashes are what finally broke my patience. Long-running research tasks, complex context synthesis, anything pushing past an hour—Claude.AI Desktop would inevitably barf and lose everything. No graceful degradation. Just gone.
So when Anthropic announced Cowork as “Claude Code for the rest of your work”? That caught my attention. This came right on the heels of /chrome going to beta for all paid users—the browser agent that lets Claude navigate websites, fill forms, and execute multi-step workflows. Two releases in quick succession that look like the first real push to close the gap between developer-focused Claude Code and everyone else.
Here’s where things get interesting: I’m on vacation this week, but I’m about to step into a new CTO role (announcement coming soon). I wanted to do proper research on my new company—reviewing their Github repos, digging through Confluence documentation, mapping out their JIRA workflow. Tasks I would have tackled with Claude Code before, but figured Cowork might handle just as well.
The verdict? Cowork is a solid step forward with one critical gap that’ll frustrate power users.
What Works
Cowork handles long-running tasks dramatically better than Claude.AI Desktop. The architecture borrowed from Claude Code makes an immediate difference.
Simon Willison’s analysis revealed that Cowork spins up a full Linux virtual machine using Apple’s VZVirtualMachine framework. Your designated folder gets mounted into a containerized environment. This explains the stability—Claude isn’t wrestling with macOS filesystem quirks directly, and crashes in the VM don’t take down your whole session.
Like Claude Code, Cowork sets up to-do lists for complex tasks, breaks work into manageable chunks, and reports progress as it goes. I kicked off a documentation analysis task and walked away for two hours. Came back to find structured notes organized by topic, saved right where I asked for them on my local filesystem.
VentureBeat’s coverage mentioned that Anthropic built Cowork in roughly a week and a half—Boris Cherny confirmed this timing on X. That recursive approach (building Cowork with Claude Code) shows in how the tool handles work. It thinks in files and folders, not conversation turns. Output lands where you can use it.
For straightforward productivity tasks—organizing a mess of screenshots into a spreadsheet, sorting a chaotic downloads folder, drafting reports from scattered notes—Cowork does the job without hand-holding.
Where It Falls Short
But here’s the thing: Cowork is no Claude Code.
My research project needed context that spanned multiple systems. Github code structure, Confluence documentation patterns, JIRA workflow history. Each piece informed the others. I expected to orchestrate this from Claude.AI—have it delegate tasks to Cowork, maybe kick off /chrome sessions to pull information from web interfaces, then synthesize everything back into my working context.
That coordination doesn’t exist yet.
What I found instead: each tool operates in isolation. The Register’s coverage captures the limitation—Cowork sessions don’t persist memory across logins, and there’s no cross-device sync. More critically for my use case, there’s no native handoff between Claude.AI conversation context and Cowork’s working memory.
I ended up building my own bridge. Created a folder that both Cowork and I could access, used artifacts as handoff files, manually copied context back and forth between sessions. Functional but clunky. Still faster than the alternative—I wouldn’t have attempted this research at all without Cowork (or Claude Code). The coordination overhead was real, but so was the output.
One TechPlanet user’s early report echoed this: “while Cowork can handle complex tasks, it sometimes lacks the nuanced understanding needed for context-specific decisions.”
Multi-session work with accumulating context? Still painful. Each Cowork session starts fresh, and you’re responsible for priming it with whatever prior work matters.
The Integration Gap
What I expected: Claude.AI as the orchestration layer. Feed it a complex research goal. Have it break the work down, delegate filesystem tasks to Cowork, dispatch browser automation to /chrome, then pull everything back together with its memory of our ongoing conversation.
What I got: Three capable tools that don’t talk to each other without significant manual intervention.
Anthropic will probably close this gap—the Connectors architecture for linking to third-party services already exists in Claude Desktop. Axios notes that Cowork can combine with Claude’s other features like the Gmail connector. But the coordination layer that would make these tools feel like one coherent system? Not there yet.
For developers comfortable with Claude Code, you already have workarounds. Shell scripts. MCP servers. Custom tooling. Cowork isn’t adding much if you’re already operating at that level.
For the productivity users Anthropic is targeting with the “not just developers” messaging? The isolation between tools will feel arbitrary. Why can’t I just tell Claude to “research this company using whatever tools you need and summarize what you find”?
The Upside
After a week of testing, I’ve settled into a workflow: Claude.AI handles memory and complex context management. It holds the ongoing conversation, maintains my research threads, keeps track of what we’ve already figured out.
Cowork is a capable executor for tasks that benefit from filesystem access and longer attention spans. The stability alone makes it worth using for anything that would have crashed Claude.AI Desktop before.
The gap is coordination. Right now, I’m the integration layer—shuffling context between tools, managing the handoffs, keeping track of what each system knows.
That’s manageable for someone who’s used to working this way. Less acceptable for the mainstream productivity users this launch targets.
The Bigger Picture
The pattern unfolding in software engineering is hitting productivity work now. Two sectors should be watching.
First: SaaS no-code solutions. Zapier, Make, all those 70% solutions that almost do what you need but never quite get there. I’ll use Cowork over any of them for complex automation, even with current limitations. If Anthropic ships cloud/desktop sync—and they’re clearly building toward it—these platforms will need to move significantly further up the stack or watch their core use cases evaporate. The friction of “almost works” loses to “actually works” every time.
Second: certain productivity roles. The pattern I wrote about Monday with pattern masters applies here. Roles heavy on information gathering, document synthesis, and routine coordination are compressing fast. The roles expanding: strategic thinking (someone still has to decide what to build), pattern recognition and synthesis (connecting dots across domains), human service work (empathy doesn’t automate), and integrators who can scale operations across systems.
The transition won’t be uniform—smaller orgs will feel it first, regulated industries slower—but the directional pressure is obvious.
The Reality of Research Preview
Fortune’s analysis positions Cowork as a competitive threat to AI productivity and automation startups. Maybe eventually. Right now, it’s what Boris Cherny said it was: “early and raw, similar to what Claude Code felt like when it first launched.”
The pricing reinforces this. $100 to $200 per month for Claude Max, which you need to access Cowork at all. That’s enterprise money for a research preview that doesn’t yet integrate with the rest of Claude’s ecosystem.
I’ll finish my company research when I get home, probably back in Claude Code where multi-session context management works better. But I’ll keep testing Cowork for the standalone tasks where its stability advantage matters.
What I’m really waiting for: seamless coordination between Claude.AI’s conversation context, Cowork’s filesystem capabilities, and /chrome’s browser automation. When those three tools talk to each other natively, we’ll have something worth getting excited about.
Until then? Cowork is a solid research preview that shows where Anthropic is heading. Just don’t expect it to replace your existing workflows quite yet.
I’m Bob Matsuoka, writing about agentic coding and AI-powered development at HyperDev. For more on AI productivity tools, read my analysis of Claude Code’s evolution to agent orchestration or my deep dive into the future of software engineering.