AI-First
Claude Cowork is Anthropic's desktop AI agent: it reads and edits your local files, controls your browser, and executes full workflows autonomously, without requiring your input at every step. Generally available since April 9, 2026 (macOS in preview since January 30, 2026, Windows since February 2026), it is included at no extra cost in every paid Claude plan.
Most people open Claude Cowork, type a question, copy the answer. That is exactly what they were doing with ChatGPT in 2023, and it is just as limited. I spent my first three weeks making the same mistake. The tool was designed for something else entirely: managing local files, browsing the web, running complete workflows while you do something else. I am convinced that the real advantage with AI is not having access to a smarter model. It is building a production system that runs outside your working hours. This guide covers exactly that: folder structure, persistent projects, skills, the /schedule command, Dispatch mode, and four workflows ready to copy based on your profile.
- 🔑 Co-Work is not Chat: setting up projects with context multiplies the power tenfold.
- ⚠️ Without folder structure and context files, every session starts from scratch.
- 💡 Scheduled tasks are the real game-changer: your computer works while you sleep.
- 🎯 A well-built Co-Work project improves with every use, unlike a simple prompt.
Chat, Co-Work, Code: three modes, three levels of output
Co-Work is the only mode of the three that acts on your environment rather than replying with text: where Chat generates an answer, Co-Work produces a deliverable.
When you install the Claude desktop app, you see three tabs. The vast majority of users stay on Chat. That is the mode that answers questions, generates text, analyzes a PDF dragged into the window. Useful for one-off questions. Not built for real work.
Code is developer territory. It writes scripts, builds products, executes technical tasks. Powerful, but you need a technical background to get anything serious out of it.
Co-Work is the only mode that functions like an employee. It accesses your local files, controls your browser, connects to Gmail, Notion, or ClickUp. It generates real Word documents with proper formatting, real Excel spreadsheets with working formulas, not text you copy-paste into a blank file. The difference is striking once you use it correctly.
Asking a question in Chat is requesting an answer. Giving a task in Co-Work is delegating a deliverable. That shift in posture changes everything.
How to install Claude Cowork (15 minutes)
The installation is faster than you might expect. Only two prerequisites: an Anthropic account on a paid plan (Pro, Max, Team, or Enterprise) and a macOS or Windows machine. Co-Work does not work in the web version: the desktop app is mandatory (official download page).
1. Download Claude Desktop: go to claude.com/download, download the installer for your OS (macOS, Windows arm64 or x64). A compatibility checker is available to confirm your machine supports Co-Work before going further.
2. Open and sign in: launch the app, sign in with your Anthropic account. The interface shows two tabs: Chat and Cowork. Click Cowork.
3. Choose your execution mode: in the Co-Work settings, two modes are available. "Ask before acting" (Co-Work requests confirmation before every action on your files), recommended when starting out. "Act without asking" (direct execution, no pause), for workflows you have already validated.
4. Set up global instructions: Settings > Cowork. Define your language, default tone, expected output format. Five lines are enough to get started.
5. Validation test: ask Co-Work to create a text file in a folder of your choice. If the file appears, local file access is working. Twenty seconds, and you will know immediately whether the setup is operational.
Tip: create a dedicated
sandbox/folder for testing before pointing Co-Work at your real files. Co-Work can read and modify your file system; experimenting on real data without a safety net is the fastest way to lose unsaved work.
Folder structure: the foundation everyone neglects
Three subfolders are all you need for an effective Co-Work setup: Context (what Claude needs to know about you), Projects (work in progress), Output (finished files). Without this structure, every session starts from scratch.
The first instinct when discovering Co-Work: point the tool at an existing folder and go. Result: chaos. Claude reads everything, understands little, produces generic output.
The right approach fits in three subfolders. Context, Projects, Output. Context holds what Claude needs to know about you. Projects stores work in progress. Output receives finished files. Simple, effective, non-negotiable.
In the Context folder, three files are enough. An about-me.md file: your role, your business, your clients, two dense paragraphs. A brand-voice.md file: your tone, your go-to phrases, what you never say. A work-preferences.md file: how you structure your documents, how Claude should handle ambiguity.
Claude reads these files automatically at the start of every new session. You never explain the same things twice.
The global instructions in the Co-Work settings reinforce this logic: default tone, output format, rules that apply to every task. Think of them as an onboarding brief for a new hire. You write it once on day one, they refer back to it indefinitely. Five minutes of initial setup, weeks of friction eliminated.
The 4 files to create before your first serious session
| File | Location | Content |
|---|---|---|
about-me.md |
Context/ |
Your role, your business, your clients, 2 dense paragraphs |
brand-voice.md |
Context/ |
Your tone, your go-to phrases, what you never say |
work-preferences.md |
Context/ |
Expected output format, how to handle ambiguity |
| Global instructions | Co-Work Settings | Cross-cutting rules for every task (language, structure, length) |
Claude loads these files automatically. You never explain the same things twice from one session to the next.
This level of structure is not unique to Co-Work. It is the same approach I apply in my AI business automation workflows: clean inputs always produce predictable outputs.
Projects: where memory becomes a competitive advantage
A Co-Work project is a persistent session tied to a dedicated folder: Claude accumulates the files it produces, the decisions made, and the corrections you requested from one session to the next. That is what turns a one-off tool into infrastructure.
A Co-Work task without a project is an isolated sprint. Useful once, then gone.
A Co-Work project is a persistent session tied to a dedicated folder. Claude accumulates the files it produces, the decisions made, the adjustments requested. When you come back a week later with new inputs, it does not start from scratch. It builds on what already exists, integrates the corrections from the previous run, and refines its outputs over time.
The difference in practice: take six client meeting summaries. Without a project, Claude processes them mechanically and produces six deliverables that are correct but lack consistency. With a configured project (column order for tracking tables, conversational tone for emails, pre-loaded client context), the deliverables are consistent, aligned with your standards, and ready to send. One practitioner described processing those six files in three minutes, for a result that would have taken two to three hours of manual work.
The more you use a Co-Work project, the more precise it becomes. That is the opposite of a tool that stagnates with every use.
The practical rule: one project per recurring workflow. Client deliverables, content production, weekly reporting, admin and finance. Each one accumulates context, each one improves over time. Eventually, it looks less like a tool and more like infrastructure.
I see this accumulation logic in every enterprise AI agent system that actually works: the value does not come from the model alone, it comes from the operational memory built around it.
Skills and scheduled tasks: when it runs without you
Skills are instruction files that specify how Co-Work should handle a type of work. Co-Work ships with a few built-in skills for Excel, PowerPoint, Word, and PDF. When you ask Claude to build a spreadsheet, it automatically loads the Excel skill and produces real formulas, not a flat table.
Custom skills go much further. A generic "content brief" skill produces a predictable, generic structure identical to what any AI would generate. The same skill trained on your audience, your tone, and your proven formats produces something that looks like what you would have written yourself. You build it once, Co-Work applies it on every run, without deviation.
Here is how these configurations compare on the criteria that matter:
| Configuration | Memory | File access | Automation | Required level |
|---|---|---|---|---|
| Chat | None | No | Zero | Beginner |
| Co-Work, standalone task | Session only | Yes | Manual | Beginner |
| Co-Work + Project | Persistent across sessions | Yes | Semi-automatic | Intermediate |
| Co-Work + Project + Skills | Persistent + specialized | Yes | High | Intermediate |
| Co-Work + Scheduling | Full | Yes | Complete | Intermediate |
One specific behavior worth knowing: AskUserQuestion. Before executing a complex task, Co-Work can generate a structured multiple-choice form to clarify your intent rather than guessing. That is the practical difference with a standard chatbot: Co-Work asks before it acts. For client deliverables or important files, the "Ask before acting" mode reinforces this behavior: every file action is confirmed by you before execution.
The computer use module in Cowork runs on Claude Sonnet 4.6, which achieves 72.5% on the OSWorld-Verified benchmark according to results published in June 2026, placing it at the estimated human level of 72 to 84% on that same benchmark. In practice, browser control remains slow (back-and-forth via screenshots): expect a minimum of 30 minutes for repetitive web tasks where a manual click takes 5 seconds.
Scheduled tasks are the next level. You define a prompt, you choose a frequency, Co-Work executes while you do something else. The /schedule command in any Cowork task opens the configuration: hourly, daily, weekly, weekdays only, or manual trigger.
Two concrete examples of scheduled tasks I use:
Morning briefing, every weekday at 8:45 AM
"Summarize unread Gmail emails from the last 16 hours. Extract actions that require a response today. Add the day's Google Calendar. Save to
briefings/YYYY-MM-DD.md."
AI watch, every Monday at 7 AM
"Find the 5 most-discussed AI topics this week on Reddit and YouTube. For each: title, why it's trending, possible content angle for my audience. Save to
watch/week-XX.md."
Weekly client report, every Friday at 5 PM
"Read the
clients/[ClientName]/outputs/folder for the week. Summarize the deliverables produced, decisions made, and open items. Generate a professional recap email draft ready to copy into Gmail. Save toclients/[ClientName]/reports/YYYY-MM-DD.md."
Content pipeline, every Sunday at 7 PM
"Read
content/ideas-pending.md. Identify the 3 most relevant topics for the coming week. For each: generate a complete brief with angle, section outline, and target audience. Save tocontent/briefs/YYYY-MM-DD-[topic].md."
Your computer must stay on and the app must remain open, but the rest happens without you.
The scheduled task is the line that separates "using AI" from "having an AI employee."
There is also Dispatch mode, which lets you send a task from your phone. The instruction arrives on your desktop, Co-Work executes it, and you receive a notification when it finishes. Research reports launched from the gym, client deliverables triggered from your car before a meeting.
To compare this approach with other automation tools like n8n, my analysis of n8n vs OpenClaw details the use cases where each solution wins.
Four workflows to copy based on your profile
The question I get most often: "Where do I start?" The honest answer: with the recurring workflow that costs you the most time each week. Here are four proven configurations, tested across different profiles.
Freelancer or consultant: One project per active client, with a specific context file covering account stakes, industry vocabulary, and expected deliverable format. Scheduled task every Friday: summary of the week's progress, open items, recap email draft. Meeting summaries become a validation step, not a writing task.
Content manager or creator: An ideas/ folder populated on the fly whenever an idea comes up. Scheduled task every Sunday evening: select the 3 most relevant topics for the week, generate a complete brief for each (angle, outline, audience). You arrive Monday morning with a ready pipeline. This is exactly what I use for ai-first.fr content production.
Executive or project manager: Gmail + Google Calendar connected via MCP connectors. Scheduled task every Monday at 8 AM: summary of unprocessed weekend emails, consolidated calendar, priority actions identified. What I call "zero-friction Monday": you enter your week without having to dig through your inbox for thirty minutes.
Sales team (Team plan): Shared project with access to CRM exports stored in a dedicated folder. After each sales meeting is transcribed and dropped into the right folder, a specialized skill extracts commitments made, updates the tracking table, and generates follow-ups. No code, no API integration.
The common requirement across all these profiles: clean inputs in the right folders. Co-Work is only as good as the structure you give it.
What this actually changes for building a serious AI system
I have said it elsewhere and I will say it again: companies do not want "AI." They want to save time and money. A properly configured Claude Cowork addresses that demand directly, with no abstraction and no vague promises.
The tool has real limitations. Scheduled tasks do not run if the computer is off. Each execution consumes quota, which eventually pushes you toward higher plans for intensive use. Memory between sessions exists through projects, but it is not infinite. The context window reaches 1 million tokens, which lets you process entire projects without losing track of previous decisions, but highly complex tasks can show error rates of 5 to 20% according to field analyses.
The real question is not technical. Does your Co-Work setup improve with every use, or does it start from scratch each time? If it is the latter, you are using a tool. If it is the former, you are building infrastructure.
The initial setup time (folders, context files, global instructions, one first skill) fits within an hour. What it saves over three weeks of use far exceeds that cost. On my own article production workflow for ai-first.fr, setting up a dedicated Co-Work project (automatic brief every Sunday, AI watch every Monday) freed up roughly three hours per week starting from the second week. That is the simplest business case there is.
A concrete point on return on investment. Someone who properly configures Co-Work for their client deliverables (dedicated project, context files, formatting skill) recovers two to three hours per week within the first few days. On a plan at 20 dollars per month, the equation is settled in just a few days of use.
Co-Work is not the answer to everything. For complex technical projects, production code, or integrations on critical enterprise systems, you need to go further. But for automating repetitive workflows with high operational value, it is one of the few tools that delivers on its promises within the first week of serious use. As long as you do not treat it like a chatbot.
Claude Cowork pricing 2026
Co-Work is included in every paid Claude plan, at no extra cost. Pro plan at $20/month: standard individual coverage, sufficient quota for regular use. Max plan starting at $100/month: high quota for multiple scheduled tasks per day and intensive projects. Team plan starting at $25/seat/month: multi-seat management with admin controls, shared skills across members, designed for teams. MCP connectors and specialized plugins are included at no additional charge in every tier (official Anthropic pricing).
Frequently asked questions about Claude Cowork
Is Claude Cowork different from Claude.ai? Yes. Claude.ai is a chatbot in the browser: you ask questions, it replies in text. Claude Cowork is a desktop application that acts on your environment: it reads and writes your local files, controls your browser, and generates real Word and Excel documents ready to send. It is the difference between an answer and a deliverable.
How much does Claude Cowork cost? Co-Work is included in the Pro plan at $20/month, the Max plan at $100/month (or $200/month depending on usage), and the Team and Enterprise plans. There is no additional cost for MCP connectors or business plugins.
Does my computer need to stay on for scheduled tasks? Yes, scheduled tasks run locally. Your machine must be on and the Claude app must be open at execution time. If you need automations without a local machine, you will need to combine Co-Work with separate cloud solutions.
What is the difference between a skill and a Co-Work project? A project is a persistent workspace with its own memory and accumulated files. A skill is a specialized instruction file (Excel, PowerPoint, content brief) that defines how Co-Work handles a type of task. The two combine: a project can load multiple skills on every run.
Does Claude Cowork replace tools like n8n or Zapier? Partially. Co-Work excels at cognitive tasks coupled with file actions (writing, analysis, document synthesis). For pure API integrations across dozens of services without a cognitive layer, n8n remains better suited. The two are complementary depending on the workflow; see my analysis of n8n vs OpenClaw.
Does Claude Cowork work on Windows or only on Mac? Both. Claude Desktop, the app that hosts Co-Work, is available on macOS and Windows. Feature parity is complete across both platforms: scheduled tasks, projects, skills, MCP connectors, and Dispatch mode all work identically.
Claude Cowork vs Microsoft Copilot Cowork: what is the difference? Microsoft announced Copilot Cowork on March 9, 2026 (available in the Frontier program since March 30, 2026), its agent integrated into Microsoft 365, built in partnership with Anthropic and powered by the same Claude models. The main difference remains architecture: Anthropic's Cowork runs locally on your machine (your files never leave your computer), while the Microsoft solution is cloud-first and integrates natively into Teams, SharePoint, and the Office 365 ecosystem. For a freelancer or small business without IT constraints, Claude Cowork is simpler to set up and operational without existing infrastructure. For a large organization already standardized on Microsoft 365, Copilot Cowork may integrate more naturally.
How long does it take to get an operational Co-Work setup? One hour is enough for a minimal working setup: installing the app (10 min), creating the three context files (20 min), configuring global instructions (10 min), and setting up a first scheduled task (20 min). Benefits are visible within the first week of regular use.
What is Dispatch mode in Claude Cowork? Dispatch is a feature launched in March 2026 that creates a channel between the Claude app on your phone and Claude Desktop on your computer. You send an instruction from your mobile, Co-Work executes it locally on your machine, and you receive a notification when it finishes. Useful for triggering tasks remotely: a research report launched from outside the office, a client deliverable kicked off before a meeting.
Can Claude Cowork connect to Notion, Slack, or Google Calendar? Yes, via MCP connectors (Model Context Protocol). These plugins let Co-Work access your online tools directly: Gmail, Google Calendar, Notion, Slack, and other common business services. They are included at no extra cost in every paid plan and are configured from the app settings, with no code required.
Vidéos YouTube
- How to Use Claude Cowork · Full Workflow Automation Guide 2026 · AI Master
- Full Tutorial: How To Automate 99% Of Your Work With Claude Cowork · Bart Slodyczka
- Master Systems Design with Claude Cowork · Blazing Zebra
- How I Automate 90% of My Social Media Content With Claude Cowork · Create Content Club