How-To Guide
How do you use Claude Cowork?
What Claude Cowork is, the plan and app it needs, how its three permission modes differ, and where the agent stops before anything reaches a platform.
By the Superdirector editorial team. Pages read November 15, 2026: Anthropic's get-started and safety articles for Cowork, the Claude documentation overview for Cowork, Superdirector's pricing page, and two competitor guides kept as labelled writing examples for structure only. Their testimonials, view counts, vendor prices and outcome claims are not adopted.
At a glance
How do you use Claude Cowork?
Start a task with an outcome instead of a prompt. Open Claude on desktop, web or mobile, select Cowork in the message box, connect the folder or app the job needs, and describe the finished thing you want. Cowork runs the session in the cloud on Anthropic's servers, reaches your own files and browser through the desktop app, and keeps working after you close the laptop. Choose the permission mode before the first task: Manual asks before each action, Auto is the default on Pro, Max, Ultra and Team plans and screens every action for safety, and Skip checks nothing. Cowork drafts, organizes and researches. It does not publish to a social platform on its own, because a post leaves Claude only through a connector whose tools you approved.
- Plan and surface: a paid Claude plan, with the desktop app required for local files, the browser and computer use.
- Mode before task: Manual for anything consequential, Auto for everyday drafting, Skip only in an environment you fully trust.
- Where it stops: Cowork reaches connected folders and connectors; publishing happens inside a connector's own tool, never from Cowork alone.
Three checks in order, and the second is the one beginners skip: the plan funds Cowork, the desktop app is what reaches local files and must stay open, and the account decides which approval options appear. Original diagram, drawn from Anthropic's get-started and safety articles read November 15, 2026.
Make it work for your brand
Find it. Remake it. Post it. Superdirector turns the viral formats in your niche into your brand's carousels, X posts and video scripts, and keeps the calendar full.
What Cowork actually is, and what it is not
Anthropic defines Cowork in one line: it uses the same agentic architecture that powers Claude Code, with no terminal required. Instead of answering one prompt at a time, Claude takes on a complex, multi-step task and executes it on your behalf. You describe an outcome, step away, and come back to formatted documents, organized files or synthesized research. The documentation overview adds the reach: it creates Excel workbooks with working formulas, PowerPoint presentations and formatted documents, and it splits complex work into parallel workstreams.
The sessions run in the cloud. Cowork runs your work in an isolated environment on Anthropic's servers, and your sessions and files live with your Claude account, which is why you can start a task on a laptop and open it from a phone. The desktop app is the bridge to your own machine: when a task needs a local file, your browser or your computer, Claude reaches it through the app. If the app is closed and the task needs something on your computer, it cannot get there.
What Cowork is not matters just as much. It is not a publishing tool. It drafts, organizes and researches, and it can call whatever connectors you have connected. The step from a finished draft to a live post belongs to the connector, and the connector's own permissions gate it. One competitor in this family sells exactly such a connector, and its plugin roundup says the quiet part plainly: none of Anthropic's official plugins publishes to a social platform. The publishing piece is always a separate tool. Keep that seam in view and the rest of the setup gets simpler.
One naming note for anyone following older tutorials. Cowork began as a distinct surface, and Anthropic has been folding it into the main Claude app so that chat and Cowork share one home. The two competitor guides read for this page both date that move to September 2026. The official help centre is less specific: the new experience rolls out to Pro, Max and Ultra plans, with Team and Enterprise following where the unified experience is turned on. If your message box no longer offers separate Chat and Cowork options, you are already on the new version and some screenshots in older guides will not match.
Before the first task: plan, app, and account
Three checks are worth doing before the first task, and they save a support ticket later. First, the plan: Cowork is available on paid plans, Pro, Max, Ultra, Team and Enterprise. The desktop app is on all paid plans; web and mobile arrive on Pro, Max, Ultra and Team, with Team still in beta, and on Enterprise when an owner enables them. Anthropic does not publish a separate Cowork price on top of the plan, so if a guide quotes one, check where the number came from.
Second, the app. Install or update Claude Desktop if the task will touch a local folder, a browser or your screen. The session runs in the cloud, but if your task uses local files, your browser, or your computer, you keep the desktop app open so Claude can reach them. This is the single most common beginner failure: the task runs, the cloud half is fine, and the local half cannot see the folder because the app was closed.
Third, the account surface. On Team and Enterprise plans, an administrator controls whether automatic approval is available to the organization and whether new tasks start in it, and the organization may require per-task approval for write-capable connector tools. If you are in that situation, the mode selector you read about may not show the option you expected. Ask your administrator before you plan a workflow around it.
Give it a folder, then give it a task
The first task should be boring and real. Create one folder for the job, put the raw material in it, and connect that folder rather than your whole drive. Anthropic's safety guidance recommends exactly this shape: Claude can only read and write files in folders you have connected, and a dedicated working folder is easier to audit than a home directory. Then describe the outcome in one sentence, for example turn these twelve receipts into one expense report sorted by date, and let Claude plan.
Read the plan before you walk away. Cowork shows its reasoning and approach as it works, and you can jump in to correct course mid-task. That is a different habit from prompting: you steer by intervening, not by writing a longer instruction up front. The competitor walkthrough that ranks for this topic models the same habit from the other direction, keeping a human as editor over every output even when the flow can publish.
Two protections are worth knowing on day one. Cowork requires explicit permission before permanently deleting any file, and Claude always asks before permanently deleting files, in any mode. And every run is isolated: the environment is created for that session, cannot reach your home or company network, and is removed when the session ends. Isolation protects your machine from the code Claude writes. It does not shrink what Claude can read through the access you granted, which is why the folder choice above is the real safety decision.
The three permission modes, and what each one lets through
Cowork ships three modes that decide when Claude pauses for your approval, and they map onto three different risk appetites. Anthropic's help centre documents all three, including the fact that the permission setting defaults to Auto on Pro, Max, Ultra, and Team plans, a detail the competitor guide states more bluntly than the official page does.
| Mode | What it does | When to choose it |
|---|---|---|
| Manual (formerly Ask before acting) | Claude pauses and asks for approval on each action; you allow or deny every request. | Sensitive files, first use of a connector, anything that sends messages, spends money or is hard to undo. |
| Auto (default on Pro, Max, Ultra and Team) | Claude keeps working and reviews each action for safety, blocking what it judges unsafe and asking you when it cannot find a safe path. | Everyday drafting, research and file organization where a bad draft costs minutes, not money. |
| Skip (formerly Act without asking) | Nothing pauses and nothing checks the actions automatically. | Only an environment you completely trust, with data you would not mind losing. |
The mode is only half the control. Each connector carries its own per-tool permission, Always allow, Needs approval or Blocked, and write tools inherently carry more risk because they can result in undesired actions. A sensible starting profile blocks the write tools you have not watched yet, runs Auto for read-only work, and switches to Manual for any task that can send, publish or pay. Anthropic's advice is to stay close to tasks with real-world consequences whatever the mode says, and no mode replaces your judgment.
There is also a mode-independent floor. Deletion always asks, in every mode. That does not make deletion safe to wave through, because a task can still be wrong about which files it wants, but it does mean the one irreversible action has a gate you cannot turn off. If a task asks to delete a batch of files you did not expect, deny it and ask Claude to list what it plans to remove first.
Scheduled tasks: work that runs without you
Scheduling is where Cowork stops feeling like a chat and starts feeling like a system, and it is also where beginners get burned. You create a task once, and scheduled tasks run in the cloud, with no device online. That is the feature: a Friday drafting job runs whether your laptop is asleep or not. It is also the risk, because nobody is watching when it runs.
The local-folder exception catches people. A scheduled task that reads a folder on your computer still needs the Claude Desktop app open, which the competitor walkthrough spells out in its own words, while a task that works only with cloud files runs on its own. If you want the hands-off version, put the raw material where the session can already see it instead of on a drive only your machine can reach.
Anthropic's safety page gives a short and unusually specific rule set for scheduled work, and it is worth following literally: start simple, with low-risk tasks like summaries before anything complex, and do not schedule tasks that access sensitive files, send messages on your behalf, make purchases, or take other actions that are difficult to undo. Review the output after each run from the Scheduled page, and pause a task you are not actively using. A scheduled task left running is a standing grant, and it should be as deliberate as any other permission.
Plugins, skills and connectors: three different things
The extension vocabulary is where most guides lose a beginner, because the three words sound interchangeable and are not. Anthropic's own documentation separates them cleanly, and the clearest one-line definition in the family comes from Blotato's own guide: a skill is one markdown file that teaches Claude a single job, a connector is a live wire to one outside tool through MCP, and a plugin bundles skills, slash commands, connectors and sub-agents into one install.
| Type | What it is | What to check first |
|---|---|---|
| Skill | One markdown file that teaches Claude a single job, such as how your brand writes or how your reports are structured. | That the job recurs weekly; a skill for a one-off adds clutter to every session. |
| Connector | A live MCP connection to an outside tool, carrying its own per-tool permissions. | Which tools it exposes and whether any of them can write, send or publish. |
| Plugin | A package of skills, slash commands, connectors and sub-agents installed together for a role. | Whether you need the whole bundle; a plugin widens what Claude can do in one install. |
Where the extensions come from is also narrower than it looks. Cowork loads the connectors, skills and plugins enabled for your claude.ai account, synced at session start, and it does not read the Claude Code directory on your machine. A skill you built in Claude Code will not appear until you add it in Customize. The official plugins themselves are free and open source but need a paid Claude plan, as the competitor roundup notes, and installing one is a scope decision: local MCP servers bundled with plugins run on your computer with the same permissions as any other program you run.
The vocabulary separates by scope, not by importance: one file for a recurring job, one live wire to an outside tool, and one install that carries several parts together. Original diagram, drawn from Anthropic's Cowork overview and the plugin guide read as a labelled writing example, both November 15, 2026.
Where Cowork stops: publishing, usage and the honest boundaries
Three limits deserve a paragraph each, because each one surprises someone every week. The first is usage. Cowork is heavier than chat, and Anthropic says so directly: multi-step tasks, like ones that run code, create files, or use your connected apps and browser, use more of your usage than a quick question. Group related work into one task, start a new conversation for unrelated work, and check usage in Settings. A task that keeps hitting the limit is usually a task that should have been three smaller tasks.
The second is the work Cowork does not run. A competitor who uses both tools reports that heavier video pipelines like Remotion did not run in Cowork when the guide was recorded, so heavy video production stayed in Claude Code. The honest summary is that Cowork covers documents, spreadsheets, research, file work and browser tasks, and Claude Code remains the place for code-heavy pipelines. The Claude Code content pipeline guide covers that half, and the connector roundup covers which connected tools are worth wiring in.
The third is trust. Cowork can read content outside the set of sources you consider safe, and the risk that follows is prompt injection: malicious instructions embedded in external content Claude reads as part of a legitimate task, which Anthropic's safety page defines in exactly those terms. Web content is a primary vector, and the official guidance is to extend internet access only to sites you trust, watch for unexpected behaviour, and remember that you remain responsible for all actions taken by Claude performed on your behalf. A browser-enabled agent with a publishing connector deserves the Manual mode, not the default.
On publishing specifically, Cowork itself still stops at the draft. A post leaves only through a connector, which is why the social-scoped version of this question lives on its own page: how to use Claude Cowork for social media explains the connector tiers and the queue, and posting to Instagram with Claude covers the account connection. This page owns the setup question; those pages own the platform question.
What to do this week
Pick one recurring job that produces a file, not a decision. Weekly reporting, receipt sorting and meeting-note synthesis are good first jobs because a wrong output is visible and cheap. Make one folder, put real material in it, connect only that folder, and run the task in Manual mode while you watch. Once two runs come back clean, switch to Auto for that job and consider a scheduled task. Keep anything that sends, publishes or pays in Manual indefinitely.
If the recurring job is social content, the layer below Cowork changes what you get. Superdirector Starter at $29 a month is the hands-on plan for people who work from Claude or ChatGPT: live social data, ideas, scripts and downloads, with you making every publishing decision. Growth at $199 a month adds scheduled research and routines and, once you turn posting on for an account, publishes eligible Instagram carousels and single X text posts by itself after one warning that names the account, the public effect and the cost. That is an account-level standing permission, granted once and revocable, and you still make the important calls about what posts, what changes and what gets paid.
The honest close is the one every good page in this family reaches. Cowork removes the busywork of assembling and organizing work. It does not remove the judgment about what should exist, and it does not publish anything by itself. A beginner who sets one folder, one mode and one real task this week will learn more than a reader who installs six plugins and schedules a task that sends messages while they sleep.