Claude Cowork

2.1 Delegate whole tasks to Claude with the Cowork tab

Claude Desktop has three modes side by side. Chat is conversation, Code is the terminal, and Cowork is where you hand over a task and get a result back.

Open Claude Desktop, switch to the Cowork tab, and decide when to use Chat, Cowork or Code

Download Claude Desktop from claude.com/download and sign in (Cowork needs a paid plan: Pro, Max, Team or Enterprise).

This is a read-only first task — Cowork looks but changes nothing — so it's a safe way to see the describe-an-outcome flow before you hand it real work. Cowork is desktop-only (macOS/Windows) and needs a paid plan.

1. Open Claude Desktop and click the Cowork tab
2. Identify the three modes – Chat, Cowork, and Code
3. Assign a read‑only task by describing it in the Tasks pane

• You'll see The Cowork tab open in Tasks mode, and Claude reporting on your Downloads folder without changing anything — your first taste of delegating a task instead of chatting.
• Takeaway Cowork is the third way to use Claude: same engine as Claude Code, no terminal, built to take a whole task off your plate — reach for it when you want a finished deliverable, not a conversation
• Check When should you reach for Cowork to delegate a whole job and get back a finished deliverable, rather than staying in Chat for a back-and-forth conversation?
• Cost Cowork is on all paid plans (Pro ~$17-20/mo, Max, Team, Enterprise) through the desktop app — there's no free tier, and it runs on macOS and Windows only. Chat in the same app has a free tier if you just want to try the interface.

2.2 Run a Cowork task from description to finished file

One Cowork loop end to end: you describe the outcome, Claude proposes a plan, you Approve, it runs, and you check the output against the source files yourself.

Do this first Delegate whole tasks to Claude with the Cowork tab

Complete a Cowork loop that produces a verified spreadsheet based on your input files

Describe the OUTCOME and the guardrails ('make a copy first', 'list anything you couldn't read'), not the click-by-click steps — Cowork plans the steps itself, and you approve the plan before it runs. Asking it to flag what it couldn't read gives you a built-in verification list.

123 Enlarge

1. 1 Show all items displays every task in list
2. 2 Review‑ready tasks awaits approval from reviewer
3. 3 Single‑use approval button confirms task with one click

Best viewed on desktop — tap Enlarge to read the numbered controls.

1. Enter the outcome description in the task box
2. Review Claude’s proposed plan and click Approve to start execution
3. When the job finishes, open the generated spreadsheet and compare a few rows with the original receipts

• You'll see A plan you approved followed by a completed spreadsheet and a short list of any items Claude could not parse
• Takeaway Describe the outcome, review the plan before it runs, then verify the result to stay in charge of what gets done
• Check Why does describing the outcome you want plus a few guardrails work better in Cowork than spelling out every click-by-click step yourself?
• Cost Multi-step tasks that read many files use more of your plan's usage than a single chat message; Max and Team plans give more headroom for long-running jobs. Always cheap to try one small task first.

2.3 Control Claude’s actions with approval modes and sandbox settings

Two guardrails, set before the work starts: an approval mode that decides when Claude must ask, and a sandbox scoped to one folder so its shell commands cannot reach the rest of the drive.

Do this first Run a Cowork task from description to finished file

Configure safe guardrails so Claude only acts where you allow and asks before irreversible steps

Stating the boundaries up front ('only inside this folder', 'never delete', 'ask before overwriting') is how you keep an autonomous agent safe. Combine this with the 'Ask before acting' approval mode for anything sensitive.

1. Choose an approval mode – select Ask before acting to require your OK for new tools or irreversible steps, or pick Act without asking only for low‑risk work
2. Scope the sandbox – point Claude at a specific project folder so its shell commands and code run inside the isolated virtual machine and cannot reach other parts of your drive
3. Watch and course‑correct – monitor Claude’s reasoning panel and intervene mid‑task to redirect or stop it if it heads the wrong way

• You'll see Claude operates solely inside the specified folder, pauses for your confirmation on risky actions, and responds when you intervene mid‑task
• Takeaway Guardrails keep agentic power safe by using ‘Ask before acting’, limiting Claude to a single folder, relying on the isolated sandbox, and interrupting whenever needed
• Check Why is scoping Cowork to a single named folder and keeping the ask-before-acting mode on the safer default for any sensitive task?
• Cost No extra cost — approval modes and the sandbox are built in. The real saving is avoiding a bad bulk action: a minute reviewing a plan beats undoing a hundred renamed files.

2.4 Use connectors to let Cowork work across Gmail, Drive, Chrome and MCP

Connectors are the services you switch on in Settings — Gmail, Drive, Slack, Chrome — so one task can read across all of them and return a single combined result.

Do this first Control Claude’s actions with approval modes and sandbox settings

Enable Cowork to access Gmail, Google Drive, Slack and your browser so a single task can gather data from multiple sources

Connectors are added in Settings then Connectors — you choose which ones to turn on. With no passwords handed to Claude, it acts through the connector you authorised. Start with read-only tasks on a connector before letting it write.

1. Add connectors in Settings, then Connectors to enable the desired services
2. Let a task span tools so it can read from Gmail, fetch a file from Drive and inspect a page in Chrome
3. Stay read-only until you trust it by keeping ‘Ask before acting’ enabled for any connector that can modify data

• You'll see Cowork pulls information from Drive, Gmail or Chrome and returns one combined result without you switching apps
• Takeaway Connectors extend Cowork beyond local files, keep them read‑only until trusted, and let one task span several tools
• Check Why should you keep a newly added connector read-only at first, and what does that let one task do that copy-pasting between separate apps could not?
• Cost Connectors themselves are part of the product; the third-party services keep their own limits. Tasks that fetch a lot from a connector run longer and use more plan usage than a local-only task.

2.5 Install a role‑specific plugin in Cowork

A plugin is a role's toolkit: skills, tool connections and slash commands installed together, so /sales:call-prep already knows which tools to use.

Do this first Use connectors to let Cowork work across Gmail, Drive, Chrome and MCP

Add a plugin built for your function so its skills, tools and slash commands are ready to use

Anthropic ships 11 open-source plugins for Cowork (Finance, Sales, Customer Support, Marketing, Legal, Data, Bio-Research and more). Each bundles skills (how your role works), connectors (your tools), slash commands, and sub-agents. Install them from inside Cowork.

1. Select Plugin library and choose the plugin that matches your function
2. Click Install to add the chosen plugin to Cowork
3. Enter any required tool connections when prompted
4. Type a slash command such as /sales:call-prep to run the pre‑configured workflow

• You'll see Cowork offers slash commands that already know your tools and execute the correct workflow for your job
• Takeaway Plugins turn Cowork into a specialist for your role by providing pre‑wired skills, connectors and commands
• Check Why does installing the plugin built for your role save you more than a connector alone, given everything a plugin bundles for the way your job works?
• Cost The plugins are free and open-source (you can read or fork them on GitHub); you still pay your Claude plan for the work, and any connected third-party tool keeps its own pricing.

2.6 Schedule a recurring Claude task

/schedule turns a task into a recurring one and lists it under Scheduled — but it produces its report only while the desktop app is open.

Do this first Install a role‑specific plugin in Cowork

Create a task that Claude runs automatically on a defined cadence

Type /schedule in a task, or open 'Scheduled' in the sidebar, to set the cadence. Important caveat: scheduled tasks only run while your computer is awake and the Claude Desktop app is open — they're not cloud jobs.

1. Type /schedule in any task to open the scheduling dialog and choose the recurrence (e.g., every Friday at 4 pm)
2. Click Scheduled in the sidebar to view and edit the newly created recurring task
3. Select a predefined template for the task so each run produces output in the same format

• You'll see A recurring entry appears in the Scheduled list and generates the same report each week while the desktop app remains open
• Takeaway Use /schedule with a fixed template for any rhythmic work, remembering the scheduler only runs when your machine is awake and Claude Desktop is active
• Check Why must you remember that a scheduled Cowork task only fires while your computer is awake and the app is open, and when does that push you toward a server-side automation instead?
• Cost Each run uses plan usage like any task; a heavy weekly job adds up over a month. The caveat to plan around is the 'computer awake + app open' requirement, not money.

2.7 Run large analyses with parallel sub‑agents

A job shaped as “do X for each of N items, then combine” is the one Cowork splits across parallel sub-agents — which is why you spot-check a few source items against the combined output.

Do this first Schedule a recurring Claude task

Create a single synthesis from many similar items by letting Claude Cowork split and run the work in parallel

Big 'do the same analysis across many items, then synthesise' tasks are where Cowork shines: for complex work it can coordinate multiple sub-agents working at once, then merge their results. You still review one final plan and one deliverable.

1. Identify a task that can be expressed as “do X for each of N items, then combine” Claude Cowork
2. Prompt Claude Cowork to break the job into subtasks and run them in parallel
3. Examine the combined output from Claude Cowork, then spot‑check a few source items to confirm accuracy

• You'll see A combined report appears quickly, having been produced by several sub‑agents working at the same time
• Takeaway Structure the prompt as ‘do X for each of N things then combine’ so Cowork can fan out the job across parallel agents and merge the results
• Check Why is a job shaped as do the same analysis across many items then combine the kind of work Cowork handles best, and why must you still spot-check a few sources after it synthesises the result?
• Cost Parallel work over many files is one of the heavier things Cowork does — it uses more plan usage than a single chat. Worth it for jobs that would take you hours; overkill for a two-file task.

2.8 Create a real deliverable using Claude Cowork

The chapter's pieces on one real deliverable of your own: the folder or connector, the plugin, the guardrails, the outcome description — then verify, and either schedule it or package it with Plugin Create.

Do this first Run large analyses with parallel sub‑agents

Deliver a complete, repeatable work output by combining connectors, plugins, guardrails and scheduling.

This is the real test: a deliverable from your own work, not a demo. Use the Finance, Sales, Support or Bio-Research plugin if one fits your role. Want to go further? Cowork's 'Plugin Create' lets you package your own workflow into a reusable plugin.

1. Choose a real, repeatable deliverable you actually produce.
2. Assemble the pieces from this course: point to the right folder or connector, install the role‑specific plugin, set guardrails and approval mode, and describe the outcome and template.
3. Verify the deliverable against your sources, then schedule it if it repeats, or use ‘Plugin Create’ to package the workflow for reuse.

• You'll see One finished, real deliverable from your own work – produced by a task you scoped, reviewed and verified – and, if it repeats, a scheduled version that keeps producing it.
• Takeaway You now have the whole loop: scope a real outcome, wire in the right connector and plugin, keep the guardrails on, verify the result, and schedule or package it — Cowork turns a recurring chore into a reviewed, repeatable deliverable
• Check Why is choosing one real recurring deliverable from your own work — rather than a demo — the right way to put the whole course together in a single Cowork task?
• Cost A full end-to-end task uses meaningfully more plan usage than a chat; that's the trade for getting back finished work. Building your own plugin is free and open-source; running it still uses your plan.