Claude Cowork is the desktop-only agentic mode of Claude that most people have never touched. While the browser version of Claude handles back-and-forth questions, Cowork works directly with files on your computer, chains multiple steps together, connects to external tools like Notion and Google Drive, and completes long-running tasks while you step away.
This claude cowork tutorial covers everything in 12 sections with 15 hands-on projects, plus the two major features that dropped after this course was recorded: Dispatch (remote task assignment from your phone) and Computer Use (Claude controlling your desktop autonomously). From installation and security to file management, data analysis, browser automation, skills, MCP connectors, and plugins. No coding required. By the end, you will know how to build real automations, set up custom skills, and package Cowork workflows as a service.
What Is Claude Cowork?
Claude Cowork is a desktop-only agentic AI tool built on the same engine as Claude Code. It runs inside the Claude desktop app and gives non-developers access to the same file-level automation, multi-step reasoning, and tool connections that developers use in Claude Code, without requiring a terminal.
The easiest way to understand Cowork is to compare it to the other two modes of Claude:
- Regular Chat: Standard back-and-forth. You upload a file, ask Claude to do something, get a result. Everything stays inside the conversation window. Great for quick one-off tasks.
- Claude Code: Terminal-based tool for developers. Features, debugging, running commands. Cowork uses the same underlying engine but is designed for non-developers who want the same power without writing code.
- Claude Cowork: Desktop app mode that reads, edits, and creates files directly in folders on your computer. You describe the outcome, step away, and come back to finished work.
Cowork is built for freelancers automating client deliverables, small business owners with repetitive document work, and content creators looking to scale their output. The key difference from Chat: Cowork works on your actual files rather than inside a conversation window.
Cowork is desktop-only and the app must stay open while a task is running. The lack of memory between sessions that existed when this course was first recorded has since been addressed through Cowork Projects, which give each project its own persistent scoped memory. Two major new capabilities, Dispatch and Computer Use, launched in March 2026 and are covered in their own sections toward the end of this guide.
See also: Claude Cowork tips and tricks
How to Install Claude Cowork
Claude Cowork requires a paid plan. This means Pro ($20/month), Max, Team, or Enterprise. It is not available on the free tier. Once you have a paid plan, installation takes about five minutes on both Mac and Windows.
Installation Steps
1. Go to claude.ai/downloads in your browser.
2. Download the Claude desktop app for your operating system (Mac or Windows).
3. Run the installer. Log into your Anthropic account when prompted.
4. Open the app. You will see two tabs at the top: Chat and Cowork. Click Cowork.
5. Windows users: you will see a banner that says ‘Additional setup needed. A Windows feature needs to be enabled to give Claude a secure workspace.’ Click Enable, confirm the Windows PowerShell prompt, and restart your computer.
After restarting, the error banner will be gone. Cowork is ready.
The interface will look familiar if you have used Claude Chat. The main difference is a folder selector on the left and a plus icon to add files, connectors, and plugins. Your model selector (Opus, Sonnet, Haiku) and extended thinking toggle are in the same location as Chat.
One important usage note: Cowork tasks are more resource-intensive than Chat. You will hit your usage limit faster because Cowork reads files, executes multi-step workflows, and may spin up sub-agents. Plan your tasks accordingly and watch your usage in Settings > Usage.
Security and Privacy in Claude Cowork
Before running any tasks, it is worth understanding what stays on your computer and what leaves it. This is more important in Cowork than in Chat because Cowork has access to your actual files.
What Stays Local and What Gets Sent
- Your files: Under normal conditions, files stay on your computer. They are not sent to Anthropic for training or storage. Cowork runs in a virtual machine on your machine.
- Your prompts and task descriptions: These are sent to Anthropic through the API, just like any Claude conversation.
- Cowork conversation history: Stored locally on your device, not on Anthropic servers. This means it does not appear in admin audit logs or compliance exports. Anthropic explicitly says not to use Cowork for regulated workloads.
What to Avoid Giving Cowork
Anthropic recommends keeping certain data away from Cowork. This includes social security numbers, passwords, financial account information, API keys, NDAs, healthcare records, and any PII. For client work, create dedicated folders for each client and copy only what is needed for a specific task rather than granting access to an entire drive.
Prompt Injection: The Real Risk
Prompt injection is when malicious instructions are hidden inside a file that Cowork is asked to process. A PDF, Word doc, or spreadsheet could contain hidden text that tells Cowork to take an action without you knowing. This is why you should only point Cowork at files you trust and always keep backups before running tasks on important files.
Approval Prompts and Deletion Protection
Before any risky action, Cowork shows you its plan and waits for your confirmation. You can approve it, redirect it, or stop it. Before deleting anything, Cowork will list what it is about to remove. Read these prompts carefully. Prompt injections can interfere with approval prompts, which is another reason to trust your source files.
The golden rule: always read your approval prompts, only work on files you fully trust, keep backups, and stay present while Cowork is running. You are the most important safety layer.
See also: Claude Cowork Projects
See also: Claude Cowork web scraping guide
See also: Claude Cowork Excel guide
See also: Claude Cowork PowerPoint guide
See also: Claude Cowork plugins guide
Need Help Building AI Automations?
Join Our AI Community
Global Instructions and Folder Instructions (claude.md)
Claude Cowork has two layers of persistent instructions: global instructions that apply to every single session, and folder instructions that are specific to a project folder. Setting both up before you start real work saves a lot of repetitive prompting.
Setting Up Global Instructions
Global instructions are preferences that Claude applies to every Cowork session automatically. To set them up: go to Settings (bottom-right corner of the app), click Cowork, and look for the Global Instructions field.
A well-built global instructions block should include:
- Your name, role, and background
- How you want Claude to work with you (example: ‘Always show me a plan before taking any action that touches files’)
- Your primary tools (BigQuery, Python, Notion, etc.)
- File safety rules (example: ‘Never delete without explicit confirmation’)
- Output format preferences
A fast way to build your global instructions: open Claude Chat, paste in your recent conversations, and ask Claude to generate a global instructions block based on your patterns. It will get you 80% of the way there in seconds. Then paste the result into Settings > Cowork > Global Instructions and save.
Folder Instructions with claude.md
Folder instructions are project-specific context stored in a file called claude.md in your project folder. Every time you open that folder in Cowork, Claude reads this file automatically. There is no setting to configure this inside Cowork. You create the file manually.
The fastest way to create a claude.md: open Claude Chat, describe your project, your data structure, and any specific conventions you use, and ask it to generate a claude.md folder instructions file. You will get a complete template to download and place in your folder.
A good claude.md for a client project folder typically includes:
- Client overview and the type of analysis being done
- Key metric definitions (for example, a rolling 30-day GMV calculation with the exact SQL logic)
- Domain context so Claude understands the business entities without needing re-explanation each session
- Tool conventions like BigQuery standard SQL, table naming formats, or safe divide patterns
- Folder structure so Claude knows where outputs should be saved
- Explicit behavior rules for ambiguous situations
File Management Projects
File management is one of the most practical things you can do with Cowork right out of the box. Cowork can read, create, rename, sort, and reorganize files in any folder you give it access to. Two key projects illustrate this well.
Project 1: Downloads Folder Cleanup
A messy downloads folder is something most people deal with but never fix. The workflow is straightforward:
1. Open a new Cowork task. Click the folder selector and choose your downloads folder (or a test copy of it).
2. When prompted, choose Allow to grant one-time access or Always Allow if this is a folder you want Cowork to always be able to access.
3. Write a prompt like: ‘Sort all files in this folder into subfolders by file type. Rename any files with generic names like ‘document1′ using the content of the file to generate a meaningful name. Create a cleanup summary.’
4. Cowork will show you its plan before touching anything. Review it and approve.
Important: always keep a backup of your downloads folder before running this. Cowork can delete files, and if it misidentifies something, you want to be able to recover it.
Project 2: Invoice Organization and Renaming
This project takes a folder of invoices in different formats and no consistent naming, and turns it into an organized structure with a subfolder per client and a standardized filename convention.
Example prompt: ‘Please create a subfolder for each customer in this folder. For each invoice file, rename it using the following format: ClientName_YYYY_MM_DD_invoice.ext. Read the content of each file to extract the client name and date if it is not in the filename.’
Cowork will read through every file to extract client names and dates, show you a full renaming plan before making any changes, ask clarifying questions about ambiguous files (for example, files with partial dates or files that share a customer name), and then execute the changes after you confirm.
The result is a clean folder structure where every invoice is categorized and named consistently. This is the kind of task that used to take an hour of manual work and now takes about two minutes of prompting plus a few minutes of execution.
Document and Report Creation
Cowork can produce complete formatted documents and save them directly to your computer. Supported output formats include PDF, DOCX, Markdown, HTML, and plain text. You can specify the format, the structure, the file name, and where to save it, all in your prompt.
Project 1: Turning Scattered Notes into a PDF Report
This project takes a folder of raw text files containing unstructured notes on five topics and produces a single formatted PDF report with an executive summary, color scheme, and section headers.
Example prompt: ‘In the scattered notes folder I have information about five card brands. Can you build this into a PDF report with proper headings and an executive summary? Use red and white colors. Save it into a new subfolder called final-report. Name the file brand-updates-Q1-2026.pdf.’
Cowork writes a Python script to generate the PDF, reads all five source files, and produces the report. If you want changes, describe them in a follow-up message. You can paste in a screenshot of the current output and say what is wrong. Cowork will fix the specific issue and regenerate the file.
Pro tip: specify that each section should start on a new page. Cowork will not always do this by default and you will end up with awkward spacing if you do not ask.
Project 2: Batch Document Generation from a CSV
This project takes a CSV of customers with columns for name, industry, plan tier, key goals, and primary challenge, and generates one customized PDF document per customer.
Example prompt: ‘In this folder you have a spreadsheet of customers. I want you to build a PDF document for each customer that covers their Q2 2026 goals. Use these columns: client name, industry, plan tier, key goals, and primary challenge. Create a new folder called Q2-2026-goals and save each PDF there. Name each file using the client name. Make the PDFs look modern with full blue colors.’
Cowork reads the CSV, installs any required dependencies (like a font package if you specify one), writes a Python script to generate all documents, and runs it. One important caveat: if you specify a font that is not easily downloadable in the execution environment, Cowork will substitute a visually similar font. It will tell you when it does this.
If the output is not quite right, upload a screenshot of what you see, describe the issue, and Cowork will fix it and regenerate all files.
Data and Spreadsheet Work
Cowork handles CSV, XLSX, and JSON files. It can extract data, clean messy spreadsheets, analyze data, and generate reports. It currently struggles with merged cells and highly irregular formatting, but it handles most real-world messy data well.
Project 1: Data Extraction from Multiple File Formats [H3]
This project extracts specific columns from five different receipt files, each in a different format (PNG, JPEG, PDF, DOCX, HTML), and combines them into a single CSV.
Example prompt: ‘I want you to organize my business dinner expenses. In this folder I have different files showing places we ate as a team. I need you to combine these together into a CSV called dinner-totals.csv. Columns to extract: restaurant, subtotal, total. Once extracted, move the original files into a subfolder called scanned. The final CSV should stay at the top level of the folder.’
Cowork reads each file regardless of format, including image-based receipts using vision capabilities. It extracts the specified columns and builds the combined CSV. Accuracy is high for clearly structured receipts, though complex or handwritten receipts may need review.
Project 2: Spreadsheet Cleanup [H3]
This project takes a messy CSV with 50 rows of concert ticket sales data and cleans it across multiple dimensions simultaneously. The before state: inconsistent band name capitalization, date formats in three different styles, prices with and without dollar signs, duplicate rows, missing values, and a revenue column with calculation errors.
A good cleanup prompt specifies each type of issue individually:
‘In this folder I want you to clean up a CSV file. There are a ton of inconsistencies. Band names are not uniform and should follow title case. Dates have different formats and should follow MM/DD/YYYY. Prices should have no dollar signs, just numbers. Remove any duplicate rows. Fill in missing values where applicable. If you do not have enough information to fill a missing value, ask me first. If there are inconsistencies with the revenue column, flag these and create a new column called redo-calculation. Rename the ticket ID column to concert ID. If you see anything else worth flagging, let me know.’
Before touching a single cell, Cowork will output a full analysis of every issue it found along with its proposed fixes. This review step is what makes it trustworthy for data work. You go through each flag, approve or adjust, and then it executes.
Project 3: Automated Weekly Report with Folder Instructions
This project uses a claude.md instructions file to define the exact format of a recurring report. The folder contains a pizza restaurant’s weekly sales data as a CSV, and the claude.md describes the brand colors, chart types, data tables, and file naming convention.
Prompt: ‘Can you build me a report based on the folder instructions and the data in the spreadsheet? Make it a PDF.’
Cowork reads the claude.md first, then reads the data file, and generates a PDF that matches the brand guidelines. When you upload a new data file the following week, you run the same prompt again and get an updated report with the same formatting. The value here is the instructions file handles all the design decisions so you never have to re-explain them.
See also: Claude Cowork Dispatch
Browser Automation with Claude in Chrome
Claude Cowork can control a Chrome browser to do research, navigate websites, fill out forms, extract data from pages, and compile information into documents. This requires installing the Claude for Chrome extension first.
Setting Up Claude in Chrome
1. Open Chrome and go to the Chrome Web Store. Search for ‘Claude.’
2. Click ‘Add to Chrome’ and accept the permissions.
3. Click Authorize to connect the extension to your Claude account.
4. Pin the extension to your Chrome toolbar.
5. Back in the Cowork desktop app, click the plus icon, go to Connectors, and you will see Claude in Chrome listed. Click Configure to review safety settings.
In Cowork Settings > Permissions, you can manage which websites Claude is approved to take actions on. You can choose to allow individual sites on a case-by-case basis or pre-approve specific trusted sites.
What Claude in Chrome Can and Cannot Do
Claude in Chrome can navigate pages, click buttons, fill forms, extract structured data from tables, and compile findings into documents. When it runs, a gold Chrome tab opens and you can watch what it does in real time.
Important restriction: financial services sites (Stripe, PayPal, etc.) are blocked by Cowork’s built-in safety filters. Healthcare sites and some large e-commerce platforms may be blocked as well. If Cowork tells you a site is blocked by safety restrictions, it is not a bug. It is intentional. If your task requires those sites, use web search instead of browser automation.
Project 1: Ultra Race Research Spreadsheet
Task: find all ultramarathon races over 100 miles happening between March and June 2026 from a specific event database website, and compile them into a spreadsheet.
Prompt: ‘Using Claude in Chrome, go to ultrasignup.com. Find me races over 100 miles taking place between March and June of 2026. Make a spreadsheet with these columns: event name, distance, date, city, format.’
Cowork opens a Chrome tab, navigates to the site, uses the search filters to narrow results, extracts the data from the results table, deduplicates entries where the same race has multiple distances listed, flags relay-format and virtual races, and saves the data as a CSV. This type of extraction works well on event listing sites that have standard table layouts.
Project 2: Competitor Pricing Comparison Document
Task: research pricing pages of three automation tools and compile a structured comparison document.
Prompt: ‘Using Claude in Chrome, find three competitors to n8n. Pull their pricing pages and compile everything into a structured comparison document covering: pricing tiers, cost per unit, key integrations, AI features, and positioning.’
Cowork opens a tab for each competitor in parallel, navigates to their pricing pages, extracts pricing structure and feature information, and compiles a Word document with a master comparison table, individual profiles, and a quick-pick guide. The document includes a note on data freshness since pricing can change.
See also: n8n AI agent
Skills in Claude Cowork
Skills are instruction packages that teach Cowork how to perform specific types of work. Think of them as reusable workflows, formatting standards, or process playbooks packaged into a markdown file. When you send a prompt, Cowork scans it and automatically activates any relevant skills it detects. You do not have to manually call them.
The difference between global instructions, folder instructions, and skills: global instructions cover general preferences for every session. Folder instructions (claude.md) cover project-specific context. Skills are repeatable workflows that activate when the task type matches.
Built-In Skills
Cowork comes with four built-in skills: Excel, PowerPoint, Word (DOCX), and PDF. These automatically activate when you ask Cowork to produce or work with those file types. To view and manage them, go to Settings > Capabilities and scroll to the bottom.
Finding Third-Party Skills
Two main resources for discovering community-built skills:
- skills.sh: Searchable marketplace with 70,000+ skills. Browse by category or search by topic. Each skill has a description, the full markdown content, and install instructions.
- skills.md: Another skills marketplace with 270,000+ skills, sortable by stars or recency.
To install a skill from either site: copy the markdown content, save it as a .md file, go to Settings > Capabilities > Add, choose ‘Upload a skill,’ and select your file. The skill will appear in your list and activate automatically when relevant.
Project: Using Built-In Skills for Data Analysis
Task: take a cleaned concert ticket sales CSV and produce both a Word document analysis and a PowerPoint presentation.
Prompt: ‘Take the spreadsheet in the folder and analyze it to produce both a Word document and a PowerPoint presentation.’
Cowork picks up the DOCX and PowerPoint skills automatically. Before executing, it asks clarifying questions: who is the audience (client, stakeholder, internal) and what should the analysis focus on. After you answer, it builds the Word document with revenue totals, band-level breakdowns, top events, and key observations, and the PowerPoint with an executive summary, revenue by band charts, fill rate analysis, and insights.
Building a Custom Skill from a Conversation
Building a skill from scratch takes a single conversation. Open a Cowork session, tell Claude what type of work you want to automate, and answer its questions about the input, desired output, and quality standards.
Example workflow for building a YouTube title skill:
Start: ‘I want to build a skill together. It is about optimizing YouTube video titles.’
Claude will ask: what is the core job of the skill (generate titles, rewrite them, or both), what factors should it optimize for, how many options should it produce, and should it be tailored to a specific channel niche or stay general.
After you answer, Claude drafts the skill and you can test it immediately in the same conversation. If the output is good, say ‘save this as a skill’ with a name. Cowork packages it and adds it to your skills library. You will also find the skill.md file saved in your active project folder.
See also: Claude Cowork Skills guide
MCP Connectors: Connecting Cowork to External Tools
MCP stands for Model Context Protocol. It is an API standard specifically designed for AI agents. In practical terms, it is what lets Cowork connect to external tools like Google Drive, Notion, Gmail, Slack, and hundreds of other services. Every connector in the plus icon menu uses MCP under the hood.
For most connectors, setup is just a sign-in. No code required. Go to the plus icon in Cowork, click Connectors, click the tool you want to connect, and follow the OAuth flow. Once connected, you can reference that tool in any Cowork prompt.
As of March 2026, the connector library has expanded significantly. Notable additions include a unified Microsoft 365 connector that gives Cowork access to Word, Excel, PowerPoint, OneDrive, and SharePoint through a single connection, plus DocuSign and FactSet for finance and legal workflows. If you work inside the Microsoft ecosystem, the M365 connector alone can replace a significant amount of manual file-moving between apps.
Project 1: Reading Documents from Google Drive
Task: use Cowork to pull five documents stored in a Google Drive folder, summarize each one, and produce a combined executive summary saved as a PDF on your computer.
Prompt: ‘Using Google Drive, grab the files in my Section 10 folder. Summarize them and give me a four-sentence executive summary. Save this as a PDF in my local project folder.’
One important limitation: Google Drive’s MCP connector can only read native Google Docs. If you upload Word documents, PDFs, or plain text files to Drive, Cowork cannot read them. You must open them in Google Docs first (right-click > Open with > Google Docs) to convert them. Once converted, the connector works correctly.
There can also be a sync delay when files are first uploaded to Drive. If Cowork reports a folder as empty right after you upload files, wait two to three minutes and ask it to recheck.
Project 2: Uploading a Spreadsheet to Notion
Task: take a 46-row cleaned CSV of concert ticket sales and upload it into a new Notion database.
Prompt: ‘Can you use the Notion create database tool to upload ticket-sales-cleaned.csv into a new database? Please link me to it so I can confirm it uploaded correctly.’
Cowork reads the CSV, creates the Notion database with the correct schema, and uploads all rows in batches of 10 (required by the Notion API rate limits). When it is done, it provides a direct link to the database.
Critical warning: the Notion connector has a known hallucination tendency when uploading large datasets. In testing, the first upload pass produced incorrect band names, wrong cities, and made-up ticket counts. After prompting Cowork to identify and fix the errors, a second pass produced correct data. Always spot-check Notion uploads against the source file before treating the data as reliable.
Plugins: Pre-Built Bundles of Skills and Connectors
Plugins are pre-built packages that combine multiple skills, connectors, and slash commands into a single installable unit. They are larger and more opinionated than individual skills. A plugin targets a specific use case: data analysis, marketing, legal, finance, sales, or productivity.
To browse and install plugins: click the plus icon in Cowork, click Add Plugins, and browse by category. Anthropic maintains a marketplace of plugins built for different industries. Click Install and the plugin becomes immediately available in your session.
Inside each plugin you get: commands (type / to trigger specific workflows), skills (the instruction files that define how to do the work), and connectors (pre-configured links to relevant external tools).
Installing and Using the Data Plugin
The Data plugin adds a full set of SQL and data exploration commands. After installing it, you get slash commands like /analyze, /build-dashboard, /explore-data, /write-query, and /validate. The plugin also bundles connectors for Snowflake, BigQuery, Databricks, Amplitude, and others.
Project: SEO Audit with the Marketing Plugin
Task: run an SEO audit on an existing blog post using the Marketing plugin.
1. Install the Marketing plugin from the plugins marketplace.
2. In a new Cowork task, type the prompt: ‘Can you run a SEO audit on [URL]?’ followed by the URL of the page you want to audit.
3. Click plus > Plugins > Marketing > SEO Audit to explicitly trigger the plugin command.
Cowork opens a browser tab, navigates to the URL, reads the page content, and generates a scored SEO report. The report includes: overall score (out of 100), missing alt text, meta description length, word count vs. competitor average, missing schema markup types, and specific recommendations with priority labels.
In testing on a real article, Cowork flagged: missing alt text on images, a 44-character meta description (too short), no featured image, 799-word article while competitors average 1,500, and no article or FAQ schema markup. These are actionable, specific fixes rather than generic advice.
Common Errors and How to Troubleshoot Them
Cowork is a relatively new product. Some things will not work perfectly out of the box. Here are the most common issues and how to handle them.
App Closes or Computer Sleeps Mid-Task
If your computer goes to sleep or the Claude desktop app closes during a task, the task stops and cannot be resumed. You have to start over. Solution: before running long tasks, go to your power settings and temporarily disable sleep mode. On Windows: System > Power Settings. On Mac: System Settings > Battery > Prevent automatic sleeping. Set it back when the task is done.
Hitting a Usage Limit Mid-Task
Cowork burns tokens faster than Chat. You can hit your usage limit in the middle of a task if the task is large enough. The task will stop with no warning. Solution: for very large tasks, break them into smaller subtasks. Run project files in batches of 5-10 rather than 50 at once. Watch your usage percentage in Settings > Usage. If you are at 80% with a large task still pending, wait for the rolling reset window.
Misinterpreted or Vague Prompts
Most Cowork failures trace back to the prompt. If Claude does something unexpected, it is usually because the prompt was ambiguous about what ‘done’ looks like. Fix: specify the exact output file names, the format (PDF, DOCX, CSV), where to save it, what columns to include, what the structure should be, and any specific formatting requirements. The more specific, the better the output.
Merged Cells and Irregular Spreadsheet Formatting
Cowork struggles with spreadsheets that have merged cells or non-standard layouts. Before working on a spreadsheet, unmerge cells and make sure row 1 contains column headers with no blank columns in between. A clean starting structure gets significantly better results.
Wrong Output Format
If Cowork produces a Markdown file when you wanted a PDF, or a Word doc when you wanted a spreadsheet, the prompt did not specify the format clearly. Always include the file extension in your prompt: ‘save this as a PDF,’ ‘create a .docx document,’ ‘output a CSV.’ If you get the wrong format and want to convert it, just ask Cowork to convert the existing file and it will.
Connectors Not Finding Files
If a Google Drive or Notion connector reports a folder as empty when you know it contains files, it is usually one of two things: (1) there is a sync delay and you need to wait 2-3 minutes, or (2) the files are in a format the connector cannot read (for Google Drive: only native Google Docs work, not uploaded PDFs or Word docs). Always do a small test with one or two files before running a large batch through a connector.
How to Sell Claude Cowork as a Service
Cowork creates a real service opportunity. Businesses have repetitive document tasks, data cleanup workflows, and recurring report generation that they currently do manually or hire for. Cowork can automate most of these in hours, not weeks, and the deliverables are packageable.
The Discovery Conversation
Start by finding one recurring task that a business does at least weekly. Document generation from a spreadsheet, expense report compilation, competitor research reports, and weekly data summaries are all high-value examples. The question to ask: ‘What is one task that takes your team 2-5 hours every week that is mostly about moving information from one format to another?’
What You Are Actually Selling
You are selling a skills file or a plugin tailored to their specific tools, file structures, naming conventions, and output formats. The deliverable is a claude.md folder instructions file, possibly a custom skill.md, and documentation on how to run the workflow. Once built, it takes a few minutes to run each week instead of hours.
The replication advantage: once you build a workflow for one client in a specific domain, roughly 80% of the work applies to the next client in the same domain. A weekly data summary workflow built for a fitness company is 80% reusable for a retail company. This is how you scale.
Cowork vs. n8n for Automation Services
Both Cowork and n8n are strong tools for selling automation services, but they target different types of tasks. Understanding the distinction helps you sell the right solution.
- n8n is best for automated recurring processes that run in the background on a schedule or trigger, with no human involvement needed. Think syncing data between apps, sending alerts, or automated data pipelines.
- Cowork is best for tasks that need AI judgment and reasoning, or that benefit from a human in the loop. Think drafting documents, summarizing research, cleaning data with ambiguous values, or organizing files based on context.
Using both tools together makes you significantly more valuable as a service provider. You can solve a much wider range of automation problems and offer end-to-end workflow coverage.
See also: n8n for beginners
Dispatch: Assign Cowork Tasks from Your Phone
Dispatch is one of the most significant features Anthropic has added to Cowork since launch. It creates a persistent, always-on conversation thread between your Claude mobile app and the Claude desktop app running on your computer. Instead of sitting at your desk to kick off a Cowork task, you can send the instruction from your phone and let your desktop run it autonomously in the background.
How Dispatch Works [H3]
When Dispatch is enabled, your Claude mobile app connects directly to your desktop Cowork session over a secure channel. You open Claude on your phone, type a task the way you normally would, and Claude routes it to your desktop app to execute. The desktop app picks it up, runs the full Cowork workflow including file access, MCP connectors, sub-agents, and skill execution, and reports the result back to your phone when it is finished.
The key word is persistent. Unlike a regular Cowork session where the context window ends when you close the app, a Dispatch-linked session holds a continuous thread. You can send follow-up instructions, ask for status updates, or refine the output from your phone without losing the prior context from earlier in the workflow. This is the same persistent agent thread model that makes Cowork significantly more capable than standard Claude Chat.
Requirements for Dispatch
Dispatch requires a Claude Pro or Max subscription. You need the Claude mobile app installed on your iOS or Android device and signed in with the same Anthropic account as your desktop app. Your computer must be awake and the Claude desktop app must be running and connected to the internet. If your computer goes to sleep mid-task, the task pauses and resumes when the machine wakes back up.
Practical Use Cases for Dispatch
- Trigger a weekly data summary from your phone on the commute in. By the time you arrive at the office the report is in your output folder.
- Send a client document for review from your phone, ask Cowork to check it against a checklist, and get a summary of issues before you sit down at your desk.
- Kick off a file organization workflow after a project wraps, without needing to be at the computer when it runs.
- Assign research tasks while away from your desk. Cowork reads the relevant URLs, pulls together a summary document, and has it ready when you return.
Dispatch works especially well alongside scheduled tasks. You can use the scheduler for things you want to run automatically, and use Dispatch for on-demand tasks you want to trigger from anywhere. Together they give you a fully remote-capable Cowork setup where your desktop is doing productive work even when you are not in front of it.
Computer Use: Claude Controls Your Desktop Autonomously
Computer Use is the most advanced capability Cowork has shipped to date. It allows Claude to directly control your desktop — opening applications, clicking buttons, filling in forms, navigating browsers, and interacting with any software visible on screen. Instead of only being able to work with files you give it directly, Claude can now operate your computer the way a human assistant would.
What Computer Use Can Do
Computer Use works by giving Claude access to your screen and input controls. Claude sees a live view of your desktop and can take actions based on what it observes. This means it can work with applications that have no API, no file export, and no MCP connector. If a human can use it by looking at the screen and clicking, Claude can use it too.
- Open a desktop application and navigate its menus without any integration required.
- Fill in and submit web forms across multi-step flows that go beyond what browser automation alone can handle.
- Copy data from one application and paste it into another, handling the kind of cross-app data movement that previously required custom scripting.
- Interact with internal company tools, legacy software, or any GUI-based system that has no modern API.
- Take and review screenshots to verify that a step completed correctly before moving to the next one.
How Computer Use Differs from Browser Automation
Cowork already had browser automation through the Claude in Chrome connector before Computer Use launched. Browser automation works inside the browser tab and is well-suited for web scraping, form filling on web pages, and navigating websites. Computer Use operates at the operating system level and is not limited to the browser. It can interact with any application on your machine, not just web pages.
In practice, the two capabilities complement each other. Browser automation is faster and more reliable for web-only tasks. Computer Use handles everything else, and steps in when the workflow moves from the browser into a desktop application or requires actions that the browser connector cannot reach.
Computer Use with Dispatch
The combination of Dispatch and Computer Use is where the full autonomous potential of Cowork becomes clear. You send a task from your phone via Dispatch. Cowork picks it up on your desktop. If the task requires opening an application, interacting with a GUI, or navigating software that has no API, Computer Use handles those steps automatically. The result is delivered back to your phone. You sent one message from your couch and your computer completed a multi-application workflow without you touching the keyboard.
Requirements and Availability
Computer Use requires a Claude Pro or Max subscription. At launch in March 2026, Computer Use is available on macOS first, with Windows support following in subsequent releases. Your computer must be awake and the Claude desktop app must be running. Because Computer Use gives Claude control of your screen, you should treat it with the same level of trust you would give any remote access tool. Only run Computer Use tasks with files, applications, and instructions that you have reviewed and trust.
Safety Considerations
Computer Use has built-in safety controls. Claude will not take irreversible actions (like emptying trash, deleting files, or making purchases) without explicit confirmation from you in the chat. It follows the same prompt injection defense rules as the rest of Cowork, which means instructions embedded in files, web pages, or application windows are flagged for your review before Claude acts on them. If a task would interact with sensitive data, financial information, or account credentials, Claude stops and asks rather than proceeding automatically.
Frequently Asked Questions
What is Claude Cowork and how is it different from Claude Chat?
Claude Cowork is the desktop-only agentic mode of Claude that works directly with files on your computer, while Claude Chat is a browser-based back-and-forth conversation tool. Cowork can read, edit, and create files in local folders, run multi-step tasks autonomously, connect to external tools via MCP connectors, and use sub-agents to work on multiple things in parallel. Chat is better for quick questions and single-turn tasks. Cowork is better for anything that involves files, multiple steps, or external tool connections.
Do I need to know how to code to use Claude Cowork?
No. Claude Cowork is designed for non-developers. All tasks are given in plain English prompts. Cowork may write Python or other code in the background to complete a task (for example, to generate a PDF), but you never need to read, edit, or understand the code. The entire interaction is through natural language.
Is my data safe in Claude Cowork?
Under normal conditions, your files stay on your computer and are not sent to Anthropic. Your prompts and task descriptions are sent through the API as with any Claude conversation. Cowork conversation history is stored locally on your device and does not appear in Anthropic servers or admin compliance exports. The main security risk is prompt injection, where malicious instructions hidden in a file can cause Cowork to take unintended actions. Only point Cowork at files you trust.
What types of files can Claude Cowork work with?
Cowork can read and produce PDF, DOCX, XLSX, CSV, JSON, Markdown, HTML, plain text, PNG, JPEG, and other common file types. It can extract data from images and scanned documents using vision capabilities. It currently struggles with spreadsheets that have merged cells or heavily irregular formatting. Files stored in Google Drive must be native Google Docs format for the Drive connector to read them.
Can Claude Cowork run tasks automatically on a schedule?
Yes. Cowork has a built-in task scheduler accessible from the Scheduled section in the left sidebar of the desktop app. You can set tasks to run daily, weekly, or at custom intervals. Scheduled tasks only run while your computer is awake and the Claude desktop app is open. You can also use the Dispatch feature to trigger and monitor tasks remotely from your phone while the main task runs on your desktop computer.
Conclusion: What to Build First
Claude Cowork is not just another AI chat tool. It is a desktop agent that closes the gap between what you can ask an AI to do and what actually happens on your computer. This guide covers the full original course plus the newest Cowork capabilities including Dispatch for remote task assignment and Computer Use for full desktop control.
The best way to get real value from this tutorial is to start with one project from a section that directly applies to your work. If you manage files, start with the downloads cleanup or invoice organization. If you do data work, start with the spreadsheet cleanup project. If you create recurring reports, set up the claude.md-driven report automation workflow.
Once you have one workflow running reliably, add a claude.md to your project folder to remove the repetitive context-setting. Then explore skills to automate the output format. Then add connectors for any external tools you use regularly.
The resources from this course, including sample folders, claude.md templates, and example skill files, are available in the free Skool community. Cowork is still early and new features are shipping regularly. Subscribe to the channel to stay up to date as the tool evolves.
