Anthropic’s Claude Cowork is the first mainstream AI agent that a lot of Mac users have actually let near their own files. It is genuinely good at what it does. But “AI agent for your Mac” is doing a lot of work in that sentence, because Cowork and Dottie are agents over two completely different things: Cowork is an agent over your documents, and Dottie is an agent over your Mac.

That distinction decides which one you want, and it is not obvious from either product page. Here is the honest version.

TL;DR

	Dottie	Claude Cowork
What it acts on	macOS itself — apps, files, calendar, mail, screen, settings	Files in folders you grant it, plus connected cloud tools
Where the work runs	Natively on your Mac	A sandboxed environment, isolated from the rest of your machine
Voice	Wake word, push-to-talk, local STT/TTS	None
Can it open or control a Mac app?	Yes — 134 native tools	No
Offline	Yes, with local models	No — requires a connection
Local models	3,800+ via MLX and Ollama	No, Claude models only
Requirements	macOS 15.4+, Apple Silicon	The Claude desktop app, plus a paid plan
Price	Free	Paid Claude plan — Pro, Max, Team, or Enterprise (pricing)

One-line answer: if the task is “reorganize this quarter’s reports and draft a summary,” use Cowork. If the task is “open Mail, find the invoice, add it to my calendar, and tell me the total out loud,” use Dottie.

What Claude Cowork actually does

Cowork is Anthropic’s agentic mode in the Claude desktop app. You point it at specific folders, and Claude gets permission to read, edit, and create files in them — so it finishes work rather than describing how you would do it. It connects to cloud tools like Google Workspace, Microsoft 365, and Slack, and it can run multi-step jobs on a schedule while the desktop app is open.

The important architectural fact is the sandbox. Cowork’s execution happens in an isolated environment that only sees the folders you explicitly grant. That is a deliberate and, frankly, a correct safety decision — you are handing an autonomous agent write access to real work product, and a blast radius of “one folder” is the right default.

It is also the constraint that defines everything Cowork cannot do. A sandbox that cannot reach the rest of your Mac cannot open Safari, cannot click a button in Finder, cannot read what is on your screen, and cannot change a system setting. Cowork is not trying to be a Mac automation tool. It is a knowledge-work tool that happens to run on a Mac.

What Dottie actually does

Dottie is a native macOS app that treats the operating system as the surface it acts on. It calls macOS APIs directly — EventKit for calendar, the Mail and Messages layers, file system operations, app launching and focus, screen OCR, system settings — through a set of 134 tools. There is no virtual machine in between, because the point is to touch the machine.

It is also voice-first, which Cowork is not at all. Wake word, push-to-talk, live transcription, barge-in interruption, spoken responses — all running locally on-device. If you have ever wanted Siri to actually complete a chained task, that is the gap Dottie aims at; we went through that comparison in detail in Siri vs Dottie.

And it runs offline. Dottie supports 3,800+ local models through MLX and Ollama, so the whole loop — speech recognition, reasoning, execution, speech synthesis — can happen on your Mac with nothing leaving it. If you have not set that up before, our guide to running local AI models on your Mac walks through it.

The real difference: sandbox versus system

Almost every practical difference between these two products falls out of one choice.

This split — sandboxed versus native — is one of four ways an AI agent for Mac can reach your system, and the tier an agent sits in predicts what it can do for you better than its feature list does.

Cowork chose isolation. The agent gets a clean, contained environment and a narrow grant of your files. You get strong safety guarantees and an agent that cannot accidentally wreck your machine. You give up the ability to automate the machine.

Dottie chose system access. The agent runs natively with real macOS permissions, gated per-tool with read-only defaults and domain restrictions. You get automation that reaches everything. You take on the responsibility of granting those permissions deliberately.

Neither is the wrong call. They are answers to different questions. The mistake is assuming a document agent will automate your desktop, or that a desktop agent is the best tool for restructuring a 200-page report.

Price, and what it actually buys

Cowork requires a paid Claude plan — Pro, Max, Team, or Enterprise. That subscription is not really “for Cowork”; it is for frontier Claude models, which are excellent and which Cowork uses for every step. You are paying for reasoning quality.

Dottie is free and its releases are open on GitHub. It is model-agnostic: run local models at zero marginal cost, or plug in your own API key for xAI, OpenAI, Anthropic, or Cerebras. If you already pay for Claude, you can point Dottie at the same Anthropic API and get frontier reasoning driving native Mac tools.

That last combination is worth saying plainly, because it is the configuration most people reading this probably want.

Where Cowork is the better tool

Being fair about this matters more than being flattering.

• Long-document work. Restructuring, summarizing, and cross-referencing large files is exactly what Cowork’s models and file-scoped design are built for.
• Cloud-tool workflows. If the job lives in Google Workspace, Microsoft 365, Slack, or Jira, Cowork’s connectors are more mature.
• Not tied to one machine. Cowork is part of the Claude desktop app and your sessions live in your Claude account; Anthropic has been extending it to run work remotely and to be driven from the Claude mobile app. Dottie is Mac-only, Apple Silicon only, and runs where you are sitting.
• Risk-averse environments. If you are not comfortable granting an agent system permissions, the sandbox is a feature, not a limitation.
• Frontier reasoning out of the box. No model configuration, no local setup.

Where Dottie is the better tool

• Anything involving an app. Open it, focus it, drive it, close it. Cowork has no path to this.
• Voice. Hands-free is not a Cowork feature at all.
• Offline and private. Local models mean the task never leaves the Mac — the reason Fello AI’s Apple Intelligence alternatives roundup called Dottie the best privacy-first Siri replacement in its comparison.
• Screen-aware tasks. OCR and visual analysis of what you are actually looking at.
• Free, and no subscription gate. Including for people who will not or cannot pay $20/month.
• Replacing macros and launchers. If you are here from Keyboard Maestro or a Raycast workflow, natural-language automation is the closer analogue.

Can you run both?

Yes, and there is no conflict. They do not compete for the same permissions or the same surface. A realistic split:

• Cowork handles the deep document sessions — the report, the deck, the research pass across a folder of PDFs.
• Dottie handles the operating system — the voice commands, the app control, the calendar and mail plumbing, the quick screen-reading tasks.

The tell for which one to reach for is simple: does the task require touching something outside a folder? If yes, it is a Dottie task.

The bottom line

Claude Cowork is an excellent agent for knowledge work, and its sandbox is a defensible design choice rather than an oversight. It is not, and does not claim to be, a Mac automation tool.

Dottie is the Mac automation tool. It is voice-first, it runs offline, it controls the operating system directly, and it costs nothing. If that is the half of the problem you have, download Dottie and try it — and if you want the wider field, our best AI assistants for Mac ranking covers 16 of them. For a straight comparison against the Claude chat app rather than Cowork, see Dottie vs Claude. For the seven-tool map, Claude alternatives for Mac. For a different agent architecture entirely, Dottie vs OpenClaw. New to the product? Start with What Is Dottie Desktop.

Frequently Asked Questions

What is Claude Cowork?

Claude Cowork is Anthropic's agentic mode in the Claude desktop app. You grant it access to specific folders and it reads, edits, and creates files to complete multi-step knowledge work, running in an isolated sandboxed environment. It requires a paid Claude plan — Pro, Max, Team, or Enterprise.

Can Claude Cowork control my Mac?

No. Cowork runs in a sandbox that only reaches the folders you explicitly grant it, so it cannot open or control Mac apps, read your screen, or change system settings. It acts on files and connected cloud tools, not on macOS itself. Dottie is the tool for system-level Mac control.

Is Dottie a Claude Cowork alternative?

For Mac automation, yes — Dottie does the system-level work Cowork's sandbox prevents: app control, voice commands, screen reading, calendar and mail actions. For long-document knowledge work, Cowork is the stronger tool. Many people run both, since they act on different surfaces.

Does Claude Cowork work offline?

No. Cowork depends on Anthropic's Claude models and requires a connection. Dottie can run entirely offline using 3,800+ local models via MLX and Ollama, so speech recognition, reasoning, and execution all stay on your Mac.

Is Claude Cowork free?

No. Cowork is included with paid Claude plans (Pro, Max, Team, Enterprise), not the free tier. Dottie is free and open on GitHub, and you can point it at your own Anthropic API key if you want Claude models driving native Mac tools.