What Cowork is — and what it isn't
Picture this: your Downloads folder has 400 files. Screenshots, invoices, meeting notes, contracts, photos that are actually Stripe receipts — all named IMG_4521.png. Sorting it manually takes hours. Sorting it by filename or extension misses the point entirely. Now imagine pointing Cowork at that folder, typing "organise this by what each file actually contains," and walking away. Thirty minutes later, every file is in a named subfolder, renamed with a date prefix, and a log file documents every decision. That's not a demo. That's Tuesday.
What Cowork actually is
Claude Cowork is Claude working directly on your local filesystem — reading, editing, and creating real files in folders you grant it access to. It plans multi-step tasks, decomposes them into subtasks, runs the work in a loop, and produces finished deliverables without you reviewing every intermediate step.
This is not "Claude with a file attachment." It is fundamentally different in three ways.
First: folder-scoped filesystem access. You don't upload one file at a time. You grant Cowork access to an entire directory tree. It can list every file, read contents, write new files, rename, move, and deduplicate across the whole structure. The folder becomes a workspace.
Second: multi-step agentic execution. Cowork doesn't respond to a prompt. It runs a loop. It plans, decomposes the goal into subtasks, executes shell commands and Python in an isolated virtual machine, and calls tools repeatedly until the work is done. You don't prompt it between steps. You approve its plan, then it runs.
Third: real file outputs. When Cowork produces an expense report, it's an actual .xlsx with working VLOOKUPs, category subtotals, and a grand total row — not a text description you paste into Excel. When it produces a deck, it's an actual .pptx with speaker notes. Finished, openable, forwardable files.
Where Cowork came from
Cowork exists because non-developers started using Claude Code — Anthropic's developer tool — for things like organising receipts and processing research notes. The terminal was the barrier. Anthropic built Cowork in roughly 10 days using Claude Code itself. It's the same agentic engine, with a UI that doesn't require you to know what a terminal is.
What Cowork is not
Not for quick questions. Agentic loops consume 5–10 times the tokens of a regular Chat conversation. Using Cowork to ask "what's the capital of France?" is wasteful. Use Chat for questions, lookups, and short drafts. Use Cowork for delegation.
Not for free-plan users. Cowork requires the Claude Desktop app and a paid subscription. Pro is the minimum; Max is recommended for sustained work.
Not for regulated workloads. Cowork is explicitly not approved for HIPAA, FedRAMP, or financial services compliance scenarios. Conversation history isn't in Anthropic's Audit Logs or Compliance API. If your data has regulatory obligations, use Anthropic's enterprise products instead.
Key Takeaways
- 1Cowork is Claude with direct access to your local filesystem — it reads, writes, renames, and organises files across an entire folder tree, not one attachment at a time.
- 2Three things separate it from Chat: folder-scoped filesystem access, multi-step agentic execution, and real file outputs (actual .xlsx, .pptx, .docx you can open and send).
- 3Cowork runs an autonomous loop — it plans, decomposes, executes, and iterates without you prompting between each step. You approve the plan; it does the work.
- 4Agentic loops consume 5–10× the tokens of Chat. Use Chat for quick questions; use Cowork for delegation tasks that produce finished deliverables.
- 5Cowork requires the Claude Desktop app and a paid account. It is not approved for HIPAA, FedRAMP, or FSI workloads.