// The guide
Claude Cowork: the AI that does the work.
Be honest about your last month with AI: how many hours did you spend executing things a chatbot had already figured out? It wrote the plan; you did the clicking, the filing, the formatting, the sending. You've been working as your AI's assistant.
Claude Cowork is Anthropic's answer to that inversion — their general-purpose agent, pitched internally as "Claude Code for the rest of your work." You give it an outcome; it plans the steps and executes them: reading and writing files, browsing the web, running code, working your mail and calendar through connected apps — and stopping to ask you when a step needs judgment.
From preview to GA in three months
The timeline matters, because it tells you how hard Anthropic is pushing this surface:
- January 12, 2026 — research preview, Max plan, macOS desktop only. Pro access four days later.
- February 2026 — plugin marketplace, admin controls, scheduled tasks.
- March 2026 — persistent agent threads across Desktop, iOS and Android; computer use preview for Pro/Max.
- April 9, 2026 — general availability on macOS and Windows.
- July 2026 — web and mobile beta (Max first), with remote sessions running on Anthropic's servers — your agent keeps working with every device off.
Availability: Pro, Max, Team and Enterprise. Nothing on Free.
The trust story: a VM between the agent and your Mac
"I don't trust an AI agent loose on my files" is the right starting position. Cowork's architecture is built as the answer, and it's worth understanding rather than taking on faith:
- On desktop, Cowork runs in an isolated virtual machine — Apple's virtualization framework on macOS, a custom Linux filesystem — not directly on your system.
- It sees only the folders you explicitly connect, mounted into the sandbox. Everything else on your disk doesn't exist for it.
- Tools are split between read and write, and you choose between auto-approve (screened) and manual approval modes.
- Deleting anything requires explicit permission. Every time.
- Against prompt injection — malicious instructions hidden in files or web pages the agent reads — there's adversarial training plus classifiers watching for suspicious instructions.
Anthropic's own framing is refreshingly unfinished: agent safety is "still an active area." Treat Cowork like a capable new hire — real work, with supervision proportional to the stakes.
What people actually use it for
Anthropic published data from 1.2 million Cowork sessions (May 2026), and it buries the "AI = coding" assumption:
- 33.4% business process work — the operational middle of jobs: organizing, processing, moving information between systems.
- 16.4% content creation — documents, decks, reports.
- 8.7% software development. Less than one session in ten.
The agent surface built by the company behind the best coding model is used overwhelmingly for non-coding work. That's the tell about where this is going.
Use cases that earn their keep
- The folder you've been avoiding: "sort these 300 files by client and year, flag duplicates, give me a summary of what's in there."
- Recurring reports: schedule a task that compiles the week's numbers into the same document format every Friday.
- Research to deliverable: "research these five vendors and build a comparison deck" — one instruction, finished file.
- Inbox-to-action: through connected apps, triage mail and draft the responses that follow a pattern.
For terminal-native work on code and files, Claude Code remains the sharper tool; Cowork is the same muscle with an office-shaped interface.
The limits nobody puts in the launch post
- It eats the shared quota. One complex Cowork session can cost as much as dozens of chat messages, from the same pool as your claude.ai chats. "Cowork rate limit reached" usually means one ambitious session, not heavy chatting.
- Local files, folders and computer use are desktop-only. Web and mobile drive remote sessions on Anthropic's infrastructure.
- Prompt injection is the #1 documented risk — be deliberate about what untrusted content you point it at.
- Enterprise gap: as of this writing, Cowork isn't covered by the Compliance API — a real blocker for regulated shops.
The missing piece: an agent that knows you
Here's what you notice after a week: every session, you re-brief it. Your clients, your formats, your standards — Cowork executes brilliantly and remembers nothing of who you are across contexts and machines. An agent with hands but no long-term memory is a gifted temp, forever on day one. Giving Claude that permanent memory — your history, your decisions, loaded before you say a word — is what we built Klyr for.
FAQ
What is Claude Cowork?
Anthropic's general-purpose AI agent: it plans and executes multi-step work — files, web, code, connected apps like mail and calendar — instead of just answering in a chat window. GA since April 2026.
Is Claude Cowork safe to use with my files?
It runs in an isolated VM on desktop and only sees folders you explicitly connect; deletions always require permission. Anthropic still labels agent safety an active area — supervise in proportion to the stakes.
What plans include Cowork?
Pro, Max, Team and Enterprise — no Free access. Usage draws from the same shared quota as chat and Claude Code.
Claude Cowork vs Claude Code — which one?
Same family, different interface. Claude Code is terminal-and-project centric; Cowork is the general-purpose version for everyday knowledge work. Anthropic's own data: only 8.7% of Cowork sessions are software dev.
Why did I hit a rate limit in Cowork?
Agent sessions are expensive — one complex session can equal dozens of chat messages, and the quota is shared across all Claude surfaces. Scope sessions tightly and check Settings → Usage.
Does Cowork keep working when my computer is off?
Yes, as of the July 2026 web/mobile beta: remote sessions and scheduled tasks run on Anthropic's servers and sync back to your devices.