Cowork by Anthropic: What an AI Desktop Agent Actually Does

Anthropic quietly shipped something that changes how Claude works — not in the cloud, not through a browser tab, but directly on your desktop. Cowork is a native macOS application that gives Claude persistent access to your screen, your files, your apps, and your running context. It’s not a chatbot wrapper. It’s closer to a colleague who can actually see what you’re looking at and help you do the work, not just talk about it.

This matters now because the gap between “AI that answers questions” and “AI that does things with you” is finally closing in a practical, shipping product. If you’ve been waiting for agentic AI to get out of demo videos and into your actual workflow, Cowork is one of the first serious attempts to deliver that — though it comes with real caveats worth understanding before you hand your desktop to a language model.

What Cowork Actually Is (And What It Isn’t)

Cowork is Anthropic’s native desktop application for macOS, built around Claude’s computer use capability — the same underlying capability that caused a stir when Anthropic first demonstrated it in late 2024. The core idea is straightforward: Claude can see your screen, interpret what’s on it, and take actions on your behalf — clicking, typing, navigating apps, reading documents, running tasks across tools without you switching contexts every thirty seconds.

What separates Cowork from simply running Claude in a browser tab is persistence and integration. The app maintains context across your working session. It can watch what you’re doing in one window while you work in another. It can be invoked with a keyboard shortcut mid-task rather than requiring you to break your flow, open a new tab, copy-paste context, and start explaining yourself from scratch.

That said, let’s be precise about what it isn’t. Cowork is not a fully autonomous agent that runs unsupervised in the background doing your job while you sleep. It’s more accurate to call it a highly capable copilot with agentic features — it can execute multi-step tasks, but Anthropic has been deliberate about keeping humans in the loop at key decision points. This is consistent with what Dario Amodei has said publicly about Anthropic’s approach to deploying powerful capabilities: the safety guardrails aren’t marketing, they’re baked into how the system behaves.

The Real Capabilities: What You Can Actually Do

Here’s where it gets practical. The things Cowork can handle that are genuinely useful in a real workday:

• Screen-aware assistance: You can highlight something on your screen — a confusing error message, a spreadsheet, a PDF — and ask Claude about it without copying anything. It reads what’s there.
• Multi-app task execution: Ask it to pull data from one place and update another. Grab a figure from a report and drop it into a slide deck. Cross-reference your calendar and draft an email accordingly.
• Long-session memory: Within a working session, Cowork maintains context about what you’ve been working on, which means you don’t restart the conversation every time you switch tasks.
• File handling: Open, read, summarize, and in some cases edit local files — documents, spreadsheets, code — directly from the app without uploading anything to a web interface.
• Coding support in-context: For developers, Cowork can see your editor, understand your current file, and suggest or write code with actual awareness of what’s already on screen — similar to what Cursor does for code specifically, but across your full desktop environment.

The computer use capability underlying all of this is still early-stage technology. Anthropic has been honest that it makes mistakes — misclicking, misreading UI elements, occasionally taking a wrong turn on a multi-step task. The current experience works best when you’re co-piloting, staying attentive, and treating it as a very fast, very capable assistant rather than a fully autonomous operator.

Cowork vs. The Rest of the Desktop AI Field

Cowork isn’t operating in a vacuum. There are several products trying to own the same space, and understanding how they differ helps you decide where to spend your time.

The honest read: Cowork is the most ambitious attempt at general desktop agency right now, but “most ambitious” doesn’t mean “most reliable.” Cursor is still the better tool if your primary use case is code. If you live in Google Docs and Gmail, Gemini’s deep Workspace integration will serve you better. Cowork earns its place when your work cuts across multiple apps and contexts — which, for most knowledge workers, is most of the time. For a direct comparison with another agentic product tackling similar territory, OpenAI Operator takes a different approach worth understanding.

Who Should Be Using This Right Now

Not everyone. Let’s be direct about the profile of someone who gets real value from Cowork at this stage versus someone who should wait.

Use it now if:

• You’re a knowledge worker — analyst, researcher, consultant, writer — who spends the day context-switching between research, documents, communication, and synthesis
• You’re comfortable with AI tools and can recognize when the model is going off the rails
• You work primarily on macOS and have a Claude Pro or Claude for Teams subscription
• You’re willing to be the human in the loop — watching, correcting, steering
• You want to compress the time between “I need to do X” and “X is done” for tasks that involve multiple tools

Wait if:

• You expect
  

  How to Get Cowork Running on Your Mac (First 30 Minutes)

  Cowork is available to Max, Team, and Enterprise plan subscribers. If you’re on a Pro plan, check your Claude account dashboard — Anthropic has been rolling access out in waves since the January 2026 research preview. Here’s exactly what to do once you have it.

  Step 1: Download and Install

  1. Go to claude.ai and sign in with a qualifying plan account.
  2. Navigate to Settings → Apps → Cowork (macOS). Download the installer.
  3. Open the .dmg, drag Cowork to Applications, and launch it. It will ask for accessibility permissions and screen recording access — both are required for it to see your screen and take actions.
  4. Sign in with the same Claude account. Cowork links to your existing conversation context and plan limits.

  Step 2: Understand the Isolated VM

  Cowork doesn’t run loose on your main Mac environment. It operates inside a local isolated virtual machine, which means the agent’s actions — opening files, navigating apps, running tasks — happen in a contained layer. Your actual system files aren’t directly exposed unless you explicitly grant access to a folder. This is worth understanding before you start, because it explains why some things (like writing directly to your Desktop) require a deliberate permission step rather than working automatically.

  Step 3: Connect Your First MCP Integration

  MCP integrations are what make Cowork useful beyond screen-watching. On first launch, go to Settings → Integrations and connect at least one: Notion, Google Drive, Linear, or whichever tool holds your actual work. Once connected, Cowork can read and write to that tool directly inside a task thread — no copy-pasting required.

  Step 4: Set Your Keyboard Shortcut and Run a Real Task

  Set a global shortcut (the default is ⌘⇧Space) so you can invoke Cowork mid-task without breaking focus. Then try this as your first real task prompt:

  “Open the last three files I worked on today, summarize what’s unfinished in each one, and draft a prioritized end-of-day checklist.”

  This hits the three things Cowork is actually good at out of the gate: reading local context, synthesizing across files, and producing something actionable. If that runs cleanly, you have a working setup.

  A Real Workflow: Contract Review in Cowork

  Contract review is one of the cleaner use cases for Cowork because the task has a clear structure: read a document, flag specific things, produce a summary. Here’s how that looks in practice using Cowork’s legal workflow plugin.

  The Setup

  • A standard vendor services agreement, PDF, sitting in a local folder you’ve granted Cowork access to.
  • Cowork open, legal plugin enabled from Settings → Plugins → Legal.
  • A second document: your company’s standard contract checklist (a simple Word doc or Notion page works).

  The Prompt Sequence

  1. Start a new Cowork thread. Drag the PDF into the thread or use the file picker to point Cowork at it.
  2. Send this prompt: “Review this vendor agreement. Flag any clauses that limit our liability beyond 1x fees paid, any auto-renewal terms shorter than 60 days notice, and any data processing terms that don’t include a DPA requirement. Use plain language for each flag.”
  3. Cowork reads the document inside the VM and returns a structured list of flagged clauses with page references and plain-English explanations. This takes roughly 45 to 90 seconds depending on document length.
  4. Follow up: “Now compare those flags against my checklist at [file path or Notion link] and tell me which checklist items this contract fails.”
  5. Final step: “Draft a short email to the vendor’s legal team listing the three changes we need before we can sign, in a professional but direct tone.”

  What Actually Happens vs. What You’d Expect

  The clause flagging is genuinely good — Claude reads dense legal language accurately and the plain-English summaries are useful for non-lawyers reviewing the output. The comparison against a second document works well when that document is structured (a numbered checklist beats a prose policy). The email draft is solid first-pass material; expect to edit the tone and add specifics, not rewrite it from scratch.

  Where it falls short: Cowork won’t catch every edge case a trained contracts lawyer would catch, and it doesn’t yet integrate with e-signature tools like DocuSign to close the loop automatically. You’re saving 60 to 90 minutes of reading time and getting a reliable first-pass review — not replacing legal counsel.

  Cowork vs. Competitors: Who Each Tool Is Actually For

  Tool	Best For	Runs Locally	File Access	Persistent Context	Weakness
  Claude Cowork	Knowledge workers doing document-heavy, multi-app tasks: legal, finance, HR, ops	Yes (macOS, isolated VM)	Local files + MCP integrations	Yes, per session thread	macOS only; Max/Team/Enterprise plans required; no real-time web browsing by default
  OpenAI Operator	Web-based task automation: form filling, bookings, browser workflows	No (cloud-based, controls a browser)	Web only	Limited	Weak on local files; lives entirely in a browser context; not useful for document-heavy work
  Perplexity Assistant	Research and real-time information retrieval; people who need current web data fast	No	No local access	No	Can’t take actions or work with your files; useful for looking things up, not doing work

  The practical answer: if your work lives primarily in local files, internal tools, and document workflows, Cowork is the only one of these three designed for that context. If you need to automate browser tasks or pull real-time web information, the other two cover ground Cowork doesn’t target.

  How to Get Cowork Running on Your Mac (First 30 Minutes)

  Cowork is available on Max, Team, and Enterprise plans. If you’re on Pro, you get the persistent agent thread on mobile and desktop but the full desktop agent requires upgrading. Here’s exactly what to do once you have access.

  Step 1: Download and Install

  1. Go to claude.ai and navigate to the Downloads section under your account settings. The Cowork app is a separate download from the Claude desktop app — don’t confuse them.
  2. Download the macOS installer. It requires macOS 13 Ventura or later. Install it like any other Mac app.
  3. On first launch, Cowork will ask for screen recording permission and accessibility access. Both are required. It runs tasks inside an isolated VM on your machine, so it is not sending live screenshots to Anthropic’s servers continuously — the VM is local.
  4. Sign in with your Anthropic account. Your plan tier is checked here. If you’re on Max, Team, or Enterprise, the full agent unlocks immediately.

  Step 2: Connect Your First MCP Integration

  Out of the box, Cowork can see your screen and read local files. The real leverage comes from MCP integrations. On first setup, you’ll see a plugin panel. Start with one integration you actually use daily — Google Drive, Notion, or your company’s legal or financial analysis plugin if your team has one configured.

  1. Click the plugin icon in the sidebar and select Add Integration.
  2. Choose from the available plugins. For most knowledge workers, start with either the file system integration (local documents) or your document cloud of choice.
  3. Authorize the connection. Cowork will confirm what it can and cannot access — read that screen carefully before clicking through.

  Step 3: Run Your First Real Task

  Don’t start with something trivial. Give it something that would normally take you 20 minutes and involves at least two steps. A good first task: “Open the Q4 report in my Downloads folder, pull out the three biggest line items by variance, and draft a four-sentence summary I can paste into the Monday standup doc.” That kind of task — read file, extract specific data, produce formatted output — is exactly what Cowork handles well and where you’ll immediately see the difference from a chat interface.

  Use the keyboard shortcut (set during onboarding, default is Command + Shift + Space) to invoke Cowork without leaving your current window. Type the task in plain language. Watch what it does in the VM panel before it writes anything to your actual files. You can pause or cancel at any confirmation step.

  A Real Workflow: Contract Review in Under 10 Minutes

  This is a practical walkthrough of using Cowork for contract review — one of the plugin-supported workflows. You do not need to be a lawyer to do this. You need a contract, a Max or Team plan, and the legal analysis plugin enabled.

  The Task

  You’ve received a vendor SaaS agreement. You need to flag anything unusual in the liability cap, indemnification, and data processing terms before sending it to legal for a proper review. Normally this takes 45 minutes of careful reading. Here’s how it goes through Cowork.

  Exact Steps

  1. Save the contract PDF to a local folder Cowork has access to — your Documents folder works if you approved file system access during setup.
  2. Invoke Cowork with Command + Shift + Space and type exactly this prompt: “Review the contract in Documents/vendor-agreement.pdf. Flag any clauses where the liability cap is below $1M, any asymmetric indemnification terms that favor the vendor, and any data processing language that conflicts with GDPR Article 28. Output a bullet list with the clause number, the issue, and your confidence level.”
  3. Cowork opens the file in the VM, reads it, and runs it through the legal analysis plugin. You’ll see it working in the agent panel — page by page if the document is long.
  4. It returns a structured list. Each flag includes the section reference, a plain-English description of the issue, and a confidence level (High / Medium / Low). Low confidence flags are things it spotted but isn’t sure about — those are the ones to send to your lawyer first.
  5. Copy the output directly into your pre-legal-review notes or ask Cowork to draft a short email summarizing the top three flags for whoever owns vendor contracts at your company.

  What This Actually Saves

  The 45-minute manual read becomes a 6-minute review of Cowork’s output, where your job shifts to judgment calls rather than extraction. You’re not replacing legal review — you’re arriving at that review with the specific questions already isolated. That’s the pattern worth understanding across every Cowork workflow: it handles the surface area, you handle the decisions that require context Cowork doesn’t have.

  Cowork vs. OpenClaw vs. Perplexity Computer: Who Each Tool Is For

  Tool	Best For	Key Strength	Main Limitation	Pricing Tier
  Claude Cowork	Knowledge workers doing legal, financial, HR, or ops work with local files and multi-app tasks	Isolated VM keeps actions local; MCP plugins for structured workflows; persistent session context	macOS only; requires Max, Team, or Enterprise plan; newer and less battle-tested	Starts at Max plan (~$100/month)
  OpenClaw (OpenAI Operator)	Web-based task automation — booking, form filling, browser-heavy workflows	Strongest at navigating live websites and completing multi-step web tasks autonomously	Less suited for local file work; no local VM; weaker on document-heavy analysis	ChatGPT Pro ($200/month)
  Perplexity Computer	Research-first users who want sourced answers driving desktop actions	Real-time web search baked into every action; strong for market research and fact-heavy tasks	Agent capabilities are shallower than Cowork or Operator; less useful for internal document work	Perplexity Pro ($20/month)

  The short version: if your work lives in local documents, internal tools, and structured professional workflows, Cowork is the right tool. If your work is browser-based and you need an agent that can navigate live sites on your behalf, Operator handles that better. If you need research synthesis with citations attached to every claim, Perplexity is faster and cheaper for that specific job. Most people doing serious knowledge work will end up with two of the three running at different points in their day.