Claude Cowork, from someone who runs it every day

Cowork is the part of the Claude app where you hand over a job instead of asking a question. Chat gives you an answer you then have to do something with. Cowork goes and does the something, inside your own files and your own tools, and comes back with the finished thing for you to check.

The practical difference is that the output is a file in the right folder rather than text in a window. No copying out, no pasting in. That sounds small and it is the whole thing.

Three properties are worth knowing before you try it:

• It works where the work is. A folder you point it at, and services you connect: Drive, Slack, Microsoft 365, anything with an MCP server.
• It shows every step. The files it opened, the tools it called, the choices it made. You can follow along from anywhere and redirect it.
• It keeps going when you stop. Close the laptop and the task continues. Put it on a schedule and it runs unattended.

This is the question I get most, and the answer is simpler than people expect: They are the same intelligence pointed at different jobs, and one paid plan gives you both, so it is not a purchase decision.

In my own week it splits cleanly. Anything that produces a repo or a running system goes to Claude Code, and there are Guides for most of those here. Anything that produces a document, a report or a browser action goes to cowork.

The honest version: start on Pro. The reason to go to Max is usage, not features, and you will know you need it because you will hit the limit in the middle of something rather than wondering in advance. The full breakdown, including what happens when you run out, is on The pricing page.

1. 1Get a plan that includes it

   Cowork is on Pro ($20 a month, or $17 if you pay for the year), Max (from $100) and Team. It is not on the free plan. The same Pro plan also includes Claude Code, so you are not choosing between them.

2. 2Install the desktop app

   claude.com/download has builds for macos, windows, windows on arm64, chromeos and linux. web and mobile exist too and are in beta, which matters more than it sounds: a task you start at your desk can be checked from your phone on the way somewhere.

   Sign in with the same account as your plan. Then look for cowork next to chat inside the app. It is not a separate download.

3. 3Give it a folder and a connector

   This is the step that decides whether cowork is useful to you or a novelty. Chat can only see what you paste. Cowork works inside a folder you choose and inside the services you connect, so the output is a file in the right place rather than text you have to move.

   Under settings, connectors, you can add Google Drive, Slack, microsoft 365, and anything that exposes an MCP server. For my own work the useful ones are drive, for where the writing lives, and Apify, for where the data comes from.

4. 4Write the task as an outcome

   The instinct from years of using software is to describe the steps. Resist it. Describe the thing you want to be holding at the end, where it should end up, and what would make it wrong.

   The difference in practice is large. "open the sheet, filter column C, copy the rows" gets you a brittle run that breaks when the sheet moves. "pull last week's signups, compare them to the week before, flag anything that moved more than ten percent, save it to Marketing/Weekly" gets you the readout, and gets it again next week.

5. 5Read the trace before you trust it

   It shows its work: the files it opened, the tools it called, the decisions it made. On the first few runs of any new task, read that. Not because it is usually wrong, but because that is how you find out what it assumed, and the assumption is the part you need to write into the prompt for next time.

6. 6Put it on a schedule once it is boring

   When a task has run correctly twice, schedule it. Daily, weekly, monthly. That is the moment the thing stops being an assistant and starts being a member of staff who does the monday readout before you are awake.

This shipped on 26 august 2026 and it changes the setup instructions in every cowork tutorial written before it, including my own video.

When a task needs a web page, cowork now opens A browser inside its own side panel instead of taking over yours. There is nothing to install. It is separate from your tabs and your logins, so it cannot wander into your email while you are not looking. You sign in once per session, or import cookies to stay signed in: chrome, edge and firefox on macos, firefox on windows and linux.

It is available in the desktop app on Pro, Max and Team. Enterprise admins turn it on under organization settings.

Anthropic's own phrasing for this is "say what, not how", and it is the single thing that separates people who get value out of cowork from people who bounce off it.

A task that works has four parts:

1. The deliverable. What you want to be holding at the end, in what format.
2. The sources. Which folder, which sheet, which connector. Name them.
3. The judgement. What counts as notable, what to flag, what to ignore.
4. Where it goes. The folder to save into.

Here is one of mine, unedited:

every monday, build last week's content readout.

sources: the youtube analytics export in Marketing/YT,
and the posting log in Marketing/Channel-tracker.

compare every number to the week before.
flag anything that moved more than 15% either way,
and say in one line what you think caused it.

one page. save to Marketing/Weekly with the date in the filename.
if a source is missing, say so at the top instead of guessing.

That last line matters more than the rest put together. Without it, a missing file becomes a confident number that is wrong, and you find out three weeks later.

In rough order of how much time each one gives back. All of them are recurring, which is the pattern: cowork earns its keep on the job you do every week, not the one you do once.

1. The monday readout. Pull the week's numbers from their sources, compare to last week, flag movement, save the doc. Scheduled, so it is done before I open the laptop.
2. Prospect research. Read a person's recent posts and write down what they are actually working on. The Full outreach system is built on this one.
3. First-pass review of a folder. Point it at twelve documents and a checklist, get twelve memos back saying where each one departs from the checklist and by how much.
4. Reconciling exports. Four regional CSV files into one summary tab with the variances flagged, plus a note on what did not reconcile.
5. Call notes into a deal memo. A quarter of notes, out comes where each deal stands, what the customer said was blocking it, and the objections that came up more than once.
6. Competitor sweeps. Open ten sites, note what changed on their pricing and landing pages since last month. This is the one the built-in browser made easy.
7. Turning a transcript into the thing it should be. An hour of talking becomes a brief, an outline and a first draft, in the folder where the writing lives.
8. The admin I keep not doing. Renaming and filing, chasing a number across three systems, filling the same form for the eleventh time.

Three features that turn a single useful run into something that runs your week.

Sub agents are specialised agents that take a piece of the task end to end. The practical effect is parallelism: while one drafts, another researches and a third organises. If a run feels like a queue, this is the thing you are missing.

Schedules run a task on a cadence, unattended. Daily, weekly, monthly. This is where the compounding is. A task you have to remember to trigger is a task you will stop doing in three weeks.

Plugins bundle skills, connectors and sub agents into one install, so cowork shows up already knowing a domain rather than being taught it every time. There is a marketplace, and organisations can run private ones.

The parts nobody puts in the launch post, from actually running it:

• It is not fast. A browser task is look, decide, click, look again. That is fine when you walk away and infuriating when you watch.
• Vague goals produce confident nonsense. The failure mode is not refusing, it is inventing something plausible. Name your sources and tell it to say when one is missing.
• The first run of anything needs reading. Budget for that. By the third run you can stop.
• Usage limits are real. A long unattended task on Pro can hit them mid-run. That is the actual reason to move to Max, and it is the only reason.
• Anything that automates a platform you do not own is your risk, not Anthropic's. Automating LinkedIn is against LinkedIn's terms, and using Claude to work around another platform's terms is against Anthropic's usage policy too. Automate the research, send the message yourself.

If you want the terminal half of this, start with The Claude Code guide, and The pricing page has the full cost picture including what the API alternative works out at. The cowork and claude code split is the only decision you have to make, and both are in the same $20.