Claude Coworker: Browser Control, Skills, and Scheduled Tasks

Introduction

Claude Coworker is the part of the Claude Desktop app that quietly turns your laptop into an autonomous workstation. It can reorganize your files, drive your browser, run scheduled jobs at 10 a.m., and dispatch work from your phone — all while you keep working in another tab. People who’ve only used Claude Chat tend to underestimate just how much heavier this side of the product is.

This guide is a hands-on tour through everything Coworker can do, what makes it different from Claude Code, the permission model that decides whether it touches your filesystem, and the four building blocks you should set up first: projects, the Chrome extension, skills, and scheduled tasks. If you’re paying for Pro or Max, this is the workflow you actually paid for.

📚 Table of contents

🧑‍💻 What Claude Coworker actually is

Coworker is the middle tab in the Claude Desktop app — the one that sits between Chat and Code. It’s an agentic Claude that runs on your machine, with permission to touch your files, drive your browser, and chain together long sequences of actions. Where Chat answers questions and Code writes software, Coworker is the one that actually does things on your computer.

⚖️ Coworker vs Code — who each one is for

Functionally, Coworker and Claude Code are siblings. Both can read files, run commands, and chain agentic steps. The real split is who they’re designed for and how much hand-holding they offer.

👉 Coworker also spins up its own virtual machine for certain operations. Code skips that and runs against your shell directly. If you’re unsure which one to use, start with Coworker — you can always graduate to Code later.

💳 Pricing, plans, and the “laptop must be on” rule

Coworker isn’t in the free tier. You need at least the $20/month Pro plan, and most heavy users sit on Max at $100–$200/month. Max is where the parallel sub-agents and long-running browser sessions stop hitting rate limits.

One workaround: the new dispatch feature lets you trigger Coworker from the Claude mobile app, as long as the desktop app is open back at home. It’s a soft version of remote control — useful, but not a substitute for true server-side scheduling.

🚀 The Coworker tab and your first task

Sign into the desktop app and you’ll see three tabs: Chat, Coworker, Code. Open Coworker and you get a control panel with a few things on the left rail — tasks, scheduled jobs, ideas, customize (skills and MCP), and recent conversations.

To start a new task:

1. Pick a model. Opus 4.6 is the most capable; Sonnet 4.6 burns usage far slower for everyday work.
2. Select a folder. This is the only directory Coworker can touch in this session.
3. Type the task. As Claude runs, you see a progress bar at the top with each tool call.

For a first run, try something low-stakes: “reorganize my desktop and colour-code the folders by category.” Coworker will scan, ask a clarifying question or two, and then start applying Finder labels and grouping files. It’s a good way to see how the agentic loop feels before you point it at anything you care about.

💡 Coworker uses significantly more credits than Chat because every tool call costs tokens. Default to Sonnet for routine work, escalate to Opus only when the task needs the bigger brain.

🔐 Folder permissions — the part that bites people

Coworker can only modify what you explicitly let it modify. When you pick a folder for the session, Claude can touch that folder and anything nested inside it — create, edit, move, delete. Outside that scope, it’s blind.

For high-stakes work, drop the relevant files into a fresh folder, point Coworker at that, and let it work in isolation. If something goes wrong, the blast radius is contained.

🗂️ Projects and persistent memory

By default, every Coworker task is amnesiac. Spin up a new session and ask “what did I have you do yesterday?” and you’ll get a polite blank stare. This is different from Claude Chat, which builds up cross-conversation memory automatically.

The fix is Projects. Hit the plus button next to “Projects” in the left rail, give it a name, and every conversation you start inside it shares a memory tab on the right. Even when the conversation gets long, Claude keeps the project’s cumulative context loaded.

🌐 Driving the browser with the Chrome extension

The browser is where Coworker starts to feel like a real coworker. Install the Claude for Chrome extension, sign in once, and Coworker can open tabs, click links, fill forms, read pages, and run multi-step browser flows on your behalf.

Setup in three steps

1. Search for “Claude for Chrome” in the Chrome Web Store and click Add to Chrome.
2. Pin the extension and sign into the same Claude account you use in the desktop app.
3. In Coworker, open the connectors panel and confirm Claude for Chrome is enabled.

Now you can ask things like:

• “Open YouTube and tell me the top two recommended videos on my home page.”
• “Click into the first video and report views, likes, and the sentiment of the top 10 comments.”
• “Search Google Flights for direct or one-stop business class flights from Miami to Dubai in the next six months under $3,000.”

For something like a flight search, Coworker will fan out into parallel sub-agents — multiple Claude instances each handling a different date range or site at the same time — and then aggregate the results into a single report. You can ask it to run things in parallel explicitly, but it usually decides on its own.

🧩 Skills: turning a one-off prompt into a reusable workflow

A skill is a Markdown file that tells Claude how to do a specific task. It can reference scripts and other resources, but at its core it’s just structured instructions: here are the steps, here are the inputs, here is the success criteria.

Good candidates for a skill:

• Anything you’d hand to an employee with a written runbook.
• Workflows you repeat weekly or daily.
• Multi-step processes where consistency matters more than creativity.
• Examples: triage incoming email, file receipts to Drive, run a research briefing, refresh a weekly report.

The build-test-save loop

1. Run the task interactively in Coworker once. Iterate on the prompt until the output is right.
2. Ask Claude to wrap the workflow into a skill, parameterised with the inputs you want to vary.
3. Let Claude test the skill itself, refine, and copy it to your skills library.
4. Next time, invoke the skill with /skill-name and pass the new inputs.

For a flight-search skill, the natural parameterisation is “source city, destination city, date range, max price, class.” The skill then asks you for those inputs when invoked, runs the same search routine every time, and produces a consistent report. You can scope skills globally or only to a specific project — project scoping is the right default for anything that depends on a particular workflow or team.

⏰ Scheduled tasks and dispatch from mobile

Skills become genuinely useful when you point a schedule at them. In Coworker, open Schedule, click New scheduled task, drop in the skill, the parameters, and a cadence (hourly, daily, weekly), and the job will fire on your local time.

You don’t even have to create the schedule by hand — just tell Coworker “add a scheduled task that runs my flight-search skill from Dubai to Miami under $4,000 every day at 10 a.m.” and it’ll wire up the cron itself.

👉 Schedules respect your machine’s timezone, not Anthropic’s servers. If you travel, double check that the cron still fires when you expect.

🔌 Connectors: the safer alternative to browser control

The Chrome extension is dramatic, but for most production workflows it’s also overkill. Connectors let Coworker hit the underlying APIs directly, which is faster, more reliable, and easier to permission.

Useful connectors to add first:

• Gmail — read and reply without opening a browser tab.
• Google Drive — file expense receipts, fetch docs.
• Calendar — schedule, reschedule, summarize the day.
• Notion — read and update workspace pages.
• GitHub — open PRs, scan issues, run repo searches.

Inside each connector, you can disable specific tools. A common pattern is to allow read-only access everywhere and only enable writes where you actively want automation. That way an over-eager Claude can’t accidentally email your boss or delete a Drive folder.

🚫 Common mistakes & pro tips

🎯 Conclusion

Coworker is the part of the Claude ecosystem that most casual users never touch — and it’s where the real productivity unlock lives. Get the basics right: a Pro or Max plan, the Chrome extension installed, a couple of connectors wired up, a project for each recurring workflow, and two or three skills that turn your most-repeated tasks into one-line invocations.

From there, scheduled jobs and dispatch take over. Your laptop becomes a quiet little agent that does the daily flight scan, files the receipts, drafts the weekly research briefing, and tells you about interesting commits before you’ve poured coffee. The work isn’t in writing more prompts — it’s in setting up the workflow once and letting it run.

Explore More on DevShelf

• →
  Claude AI — Tool Profile

  In-depth review of the Claude model family — the engine behind Coworker's browser, file, and sub-agent capabilities.

• →
  Claude Ecosystem Guide: Chat, Code, and Desktop

  How Coworker fits alongside Claude Chat, Claude Code, and Claude Desktop — and when to reach for each.