Claude Cowork is the second tab of the Claude desktop app — Claude with file access, a plan-then-act loop, and outputs back to your folder. Mid-market guide.
TL;DR. Cowork is the second tab inside the Claude desktop app. It is Claude with hands: it reads the files in a folder you grant it, edits them, creates new ones, and runs multi-step work without you having to babysit each step. It is not a chatbot, not a coding tool, and not a cloud service that scrapes your drive — it works on the laptop, on the folder, on the work.
Cowork is Claude operating directly on local files, with a plan-then-act loop the user approves. It runs inside the Claude desktop app on macOS and Windows, under any paid Claude subscription.
For a 200-person company, the most useful mental model is this: Cowork behaves like a junior analyst who only works inside one folder, can use Office and Google formats, can read PDFs and images, and never copies your data to a personal account. It is fast, it is patient, and it has the same blind spots a junior analyst has — which is why governance and a workspace pattern matter from week one, not week six.
Cowork left research preview and reached general availability on April 9, 2026. Before GA, it was a beta we put in front of our most adventurous clients with a long list of caveats. After GA, the caveat list shrank substantially.
Three capabilities arrived between late 2025 and the GA milestone that change the deployment math:
The honest implication for buyers: this is no longer a science project. Governance is still maturing — Anthropic is shipping changes weekly — but the core behaviour is stable enough to deploy at mid-market scale today.
The Claude desktop app gives you three tabs on the same subscription:
One subscription, three tabs. The decision about which tab someone uses is a per-role question, not a per-product procurement question. We cover the split on a dedicated page — see Cowork vs Chat vs Code.
Three vignettes from the kinds of users we deploy Cowork for.
A finance manager preparing the month-end close drags fourteen invoice PDFs into the inbox folder, asks for the journal entries Excel, and gets back a workbook with a Reconciling tab flagging anything outside tolerance. The whole thing takes twenty minutes; before Cowork it took an afternoon.
A marketing operator building a competitive teardown drops three competitor PDFs into a research folder, asks for a one-page comparison written in the company's house style, and receives a .docx she sends straight to the exec team after a five-minute review.
An operations lead turns a pile of supplier emails into a price-comparison sheet by exporting the thread to a folder and asking Cowork to extract pricing per SKU, normalise units, and recommend a pick. Cowork lists assumptions, flags two suppliers whose units it could not reconcile, and ships the sheet.
These are not "look what AI can do" demos. They are work that someone was already doing, faster.
Misconceptions cost more than the tool itself, so it is worth being explicit about the boundaries:
Three points that separate "Cowork demo" from "Cowork in production at a real company":
Per-seat economics are easy. A Pro seat at $20/month covers Chat, Cowork, and Code on the same account. There is no separate Cowork SKU. Procurement is one decision, not three.
The value is in workflows, not the tool. Buying licences without designing the work is the most reliable way to waste money on Cowork. This is exactly why a third of this bible is the Rollout playbook — implementation is the product, not the model.
Governance is real and immediate. Cowork edits real files. By week two of any deployment you need a folder policy, a rollback story, and an approval-gate convention. Section 04 covers all of it. Skip those, and Cowork will eventually do something legitimate that someone wishes it had not.
A four-question self-check before you commit:
If you answered yes to the first two and have a clear answer to the last two, you are in the right zone.
We do not sell Cowork — Anthropic does. We sell the implementation. The reason is uncomfortable but consistent: mid-market wins or loses on workflow design, not on whether the tool exists. We have watched companies install Cowork on Friday and abandon it by the next quarter because nobody mapped which roles get which tab, nobody picked a pilot function, and nobody owned the folder policy.
Our default first engagement is two weeks of shadowing a function, then a four-week build. We watch the work before we touch a prompt. The full method is on the Observe → Co-design → Ship page.
Look at the files in ~/cowork-workspace/inbox.
Group them by document type (invoice, contract, slide deck, other),
make a folder for each type, move the files in,
and write a 5-line summary called inbox-summary.md.
Plan first, then ask before moving anything.
Run this with Cowork's plan-mode on. It will tell you what it intends to do before it touches a file. That single habit — read the plan, approve, watch — is the mental model for everything else in this bible.
Book a 30-minute call. We'll ask where you are, what your team needs, and which systems Cowork should touch.