Claude Cowork review and how it works with Tallyfy

Claude Cowork is Anthropic’s attempt to bring the power of Claude Code to people who don’t write code — and after spending real time with it, I’ve got opinions. Here’s how we approach workflow automation at Tallyfy.

What Claude Cowork actually is

Strip away the marketing and here’s what you’ve got: Claude Cowork gives the Claude AI model access to a folder on your computer. That’s it. That’s the product.

But that simple idea is more powerful than it sounds.

Traditional AI chatbots live in a text box. You paste something in, you get something back. Cowork breaks out of that box. It can read your files, understand their contents, create new ones, edit existing ones, and organize them. It handles spreadsheets, slide decks, reports, research documents — basically the kind of knowledge work that fills most office workers’ days.

Anthropic noticed something interesting before building Cowork. People were using Claude Code — a developer tool that runs in the terminal — for things that had nothing to do with coding. Vacation research. Building presentations. Recovering wedding photos from hard drives. Monitoring plant growth. Controlling ovens. The engineering team realized the underlying capability (an AI that can take actions on files) was too useful to lock behind a terminal.

So they built Cowork in about 10 days using Claude Code itself. Felix Rieseberg from Anthropic’s team said they “spent more time making product and architecture decisions than writing individual lines of code.” That’s either exciting or terrifying, depending on how you feel about AI building AI tools.

In our experience with workflow automation, the pattern is familiar. The tools that stick aren’t the ones with the longest feature lists. They’re the ones that remove a specific friction point that was driving everyone crazy. Cowork removes the friction between “AI can help me” and “AI can do it.”

How Cowork runs under the hood

This is where it gets genuinely interesting from a technical standpoint.

Cowork doesn’t run directly on your Mac. It runs inside a sandboxed Linux virtual machine using Apple’s Virtualization Framework. Specifically, it’s a complete Ubuntu 22.04 LTS instance running on ARM64 architecture. The VM boots in about 30 seconds and delivers near-native CPU performance.

Why does this matter? Security. The AI agent can’t touch anything on your Mac that you haven’t explicitly shared. It can’t browse your filesystem, can’t access your browser passwords, can’t snoop through your email. The only thing it sees is the folder you point it at. Network access is restricted to an allowlist — basically just package registries and Anthropic’s API. Syscalls are limited by seccomp filters. There’s even a bubblewrap sandbox layer inside the VM itself.

It’s isolation within isolation. Paranoid engineering, in the best way.

One thing that caught some users off guard — Cowork creates a roughly 10GB VM bundle on your Mac without much warning. If you’re tight on disk space, that’s worth knowing before you enable it.

The Windows version, launched in February 2026, uses Microsoft’s Host Compute System instead of Apple’s Virtualization Framework, but the principle is the same. Isolated VM. Restricted access. The AI gets hands, but they’re in a box.

What you can do with Cowork in practice

I’ve been testing this for a few weeks. Here’s what’s worked well and what hasn’t.

Where Cowork shines:

Batch file processing. Point it at a folder of 50 PDFs, tell it to extract key data points and build a summary spreadsheet. Done. This used to be a full afternoon of copy-paste tedium. Cowork handles it while you do something else.

Report generation from messy data. Give it a folder of CSVs with inconsistent formatting, ask it to normalize the data and produce a clean report. It figures out the patterns and handles the edge cases surprisingly well.

Document creation and editing. Need a slide deck based on a research document? A proposal based on a template? Meeting notes organized into action items? These are Cowork’s sweet spot.

You can queue multiple tasks and let it process them simultaneously. Anthropic describes this as feeling “much less like a back-and-forth and much more like leaving messages for a coworker.” That’s accurate. You fire off a request, go make coffee, come back to results.

Where it struggles:

Anything requiring judgment calls about your specific business context. Cowork doesn’t know your company’s approval hierarchy, your naming conventions, or your compliance requirements. It can process files brilliantly, but it can’t decide whether the output is right for your organization.

And this is the gap that keeps coming up. Agents without workflows are reasoning engines with nothing structured to reason about.

Honest limitations and criticism

I’m not going to sugarcoat this. Cowork has real problems.

Quota consumption is aggressive. A single Cowork session doing complex file operations can burn through as much quota as dozens of regular chat messages. On the Pro plan at $20/month, you’ll hit your limits fast. Anthropic warns about this, but the warning doesn’t quite convey how quickly it happens, and my guess is most Pro subscribers will feel quota-constrained within a few days of heavy use. Sessions don’t sync across devices. What you do in Cowork on your Mac stays on your Mac — no web access, no mobile, no picking up where you left off on a different machine, and for a product called “Cowork,” the collaboration story is surprisingly thin. Key features are missing. Projects, chat sharing, artifact sharing, and Memory don’t work with Cowork yet, and you can’t switch between Cowork and regular chat mid-conversation — these aren’t minor gaps but the kind of features that make the difference between a demo and a daily tool. Destructive actions are possible. Anthropic’s own documentation warns explicitly that Cowork can “take potentially destructive actions, such as deleting a file that is important to you or misinterpreting your instructions,” and they also flag prompt injection risks where malicious content in files could manipulate Claude’s behavior, so always work on copies and don’t give it access to your only copy of something important.

Mac-only at launch, Windows added later. If you’re on Linux, you’re out of luck for now.

Honestly, I expected better on the sync and collaboration front. A tool named “Cowork” that doesn’t sync across devices feels like it shipped before it was ready. The research preview label explains it, but doesn’t excuse it for people paying $100-200/month.

How Tallyfy MCP makes Cowork useful for real workflows

Here’s where my frustration with Cowork turns into something more interesting.

Cowork is brilliant at file operations. But files aren’t workflows. A spreadsheet isn’t a process. A report isn’t an approval chain. If all you do is point Cowork at folders and say “organize these,” you’re using about 10% of what AI agents can do.

The missing piece is structured workflow infrastructure. And that’s exactly what Tallyfy’s MCP server provides.

MCP — the Model Context Protocol — is a standard that lets AI models discover and use external tools. Anthropic created it, then donated it to the Linux Foundation. Every major AI platform now supports it. Tallyfy was one of the first workflow platforms to build an MCP server, and it exposes 40+ tools that any MCP-compatible AI can use.

When you connect Cowork to Tallyfy’s MCP server, the conversation changes completely. Instead of “organize my files,” you’re saying things like:

• “Create a task for Sarah to review the Q1 budget by Friday”
• “Launch the client onboarding process for Acme Corp with these kick-off details”
• “Show me all overdue tasks in the compliance workflow”
• “Build a new template for vendor approval with three review steps”

The AI isn’t just moving files around anymore. It’s managing actual business processes — tasks with owners, deadlines, conditional logic, and audit trails.

We’ve observed that operations teams get dramatically better results when AI works within defined processes rather than ad-hoc. At Tallyfy, we’ve seen this pattern repeatedly: the companies that succeed with AI automation aren’t the ones with the fanciest models — they’re the ones who defined their workflows clearly enough that an AI agent can follow them. In discussions we’ve had about AI agent integration, the same question keeps coming up: “How do I get the AI to do more than answer questions?” The answer is always the same — give it a workflow to follow, equip it with tools to act within that workflow, and give it guardrails so it doesn’t go off-script. That’s what MCP + Tallyfy does. Cowork provides the AI brain. MCP provides the communication standard. Tallyfy provides the workflow structure. None of them is the whole answer by itself.

Should you pay for Claude Cowork

Depends entirely on what you’re trying to do.

Worth it if: You spend hours each week on file processing, document creation, data organization, or report generation. Cowork will save you real time on these tasks. The $20/month Pro plan is a reasonable starting point, though you might want Max at $100/month if you’re going to use it daily.

Not worth it if: You’re looking for a workflow automation tool. Cowork alone doesn’t manage processes, track tasks, or handle approvals. It processes files. That’s valuable, but it’s not the same thing as workflow automation.

The real play: Combine Cowork with Tallyfy’s MCP server. Use Cowork’s file processing strengths for document-heavy tasks. Use MCP tools for structured workflow actions — creating tasks, launching processes, managing templates, tracking progress. Let the AI agent work within a defined process rather than freestyling.

After 10 years building workflow software at Tallyfy, here’s what I’ve learned about AI tools: they’re multipliers, not replacements. The counterintuitive part is that a great AI agent multiplied by a broken process gives you a fast-moving mess. The same agent multiplied by a well-defined workflow gives you something genuinely useful.

So before you sign up for Cowork — or any AI agent tool — ask yourself: do I have a process for this AI to follow? If the answer is no, that’s the first problem to solve. And that’s exactly what Tallyfy was built for.

Related questions

What is Claude Cowork

Claude Cowork is an AI agent feature in Anthropic’s Claude Desktop app that gives Claude access to files on your computer. It runs inside a sandboxed Linux VM using Apple’s Virtualization Framework (or Microsoft’s Host Compute System on Windows), can read, edit, and create files in shared folders, and handles tasks like document processing, data analysis, and report generation. It launched January 12, 2026 as a research preview.

Is Claude Cowork free

No. Cowork requires at least a Claude Pro subscription at $20/month. It was initially exclusive to Max subscribers ($100-200/month) but Anthropic expanded access to Pro on January 16, 2026. Be aware that Cowork tasks consume significantly more quota than regular chat — a single complex session can use as much as dozens of normal messages.

Can Claude Cowork access all my files

No. Cowork can only access folders you explicitly share with it. It runs inside an isolated virtual machine with restricted network access and limited system calls. It can’t browse your filesystem, access your browser, or reach anything outside the shared folder. That said, always work on copies of important files — Anthropic warns that Cowork can take destructive actions like deleting files if it misinterprets instructions.

How does Claude Cowork connect to Tallyfy

Through Tallyfy’s MCP server. MCP (Model Context Protocol) is a standard that lets AI models discover and use external tools. When configured, Cowork can use Tallyfy’s 40+ MCP tools to create tasks, launch processes, build templates, search workflows, and manage workflow operations — all through natural language commands instead of navigating a UI.

What is the difference between Claude Code and Claude Cowork

Claude Code runs in a terminal and is designed for software developers — it writes, edits, and debugs code, runs commands, and manages git repositories. Claude Cowork runs in the Claude Desktop app and targets knowledge workers — it processes documents, creates reports, organizes files, and handles general office tasks. Both run in sandboxed environments. Both use the same underlying Claude model. The difference is the interface and the intended audience.