Claude Cowork Cheat Sheet
Every feature, setting, and limitation — a bookmark-worthy reference for legal professionals
Claude Cowork is Anthropic's agentic desktop product — Claude running inside an isolated Linux virtual machine on your Mac or Windows computer, able to execute multi-step tasks against files and applications you explicitly grant it access to. Over the last year it has grown from a single research preview into a composable stack: an instruction layer (Global Instructions, Folder Instructions, CLAUDE.md), an execution layer (Dispatch, Recurring Tasks, Sub-Agents), and a composition layer (Projects, Skills, Plugins). This cheat sheet is built to bookmark — each concept gets a crisp definition, a practical 'Use it when' line, and a link back to the canonical Anthropic documentation [1].
What Cowork Is (and Isn't)
- +What it is: A desktop agent that runs multi-step tasks on your behalf using your local files, your installed applications, and connected tools. Think of it as Claude with a workstation of its own, sitting next to yours.
- +Where it runs: Inside an isolated Linux virtual machine on your computer. Code and shell commands never touch your host operating system.
- +What you grant: Folder-by-folder access. Claude can read and write only within directories you explicitly mount.
- +What it isn't: A cloud service you share with a team. Cowork is tied to one desktop machine, one user, one active Claude Desktop app.
How Cowork Works Under the Hood
Cowork boots a lightweight, custom Linux virtual machine each time it runs. On macOS, it uses Apple's Virtualization Framework (the same technology powering tools like UTM and Tart) — a full VM with its own kernel, not a Docker-style container. The agent lives inside Ubuntu 22.04 with a disposable 10 GB root filesystem. Anything Claude does — browsing, writing code, running shell commands, calling APIs — happens inside that VM, not on your actual operating system [10].
The Instruction Layer
Global Instructions, Folder Instructions, and CLAUDE.md
Three places you can write persistent guidance that Claude will read before every task. They compose — each layer adds to the ones above it.
Global + Folder + CLAUDE.md — How They Layer
- +Think of it as three nested envelopes:
- +Global Instructions are defaults for every session. Your identity, tone preferences, and absolute rules belong here.
- +Folder Instructions add project-specific context when you open a folder. They do not replace Global — they extend or override it for that directory.
- +CLAUDE.md is file-level persistent context that travels with the folder itself. Even if you move the folder, share it via Dropbox, or re-open it months later, the context comes back because it lives in the directory.
- +A practical pattern: Put your firm-wide style guide in Global, put per-client rules in Folder Instructions, and drop a CLAUDE.md inside each matter folder with matter-specific facts (parties, jurisdictions, key dates). Claude reads all three before starting work.
Deep Dive Coming: CLAUDE.md Files
- +CLAUDE.md deserves its own article — where to put them, how long they should be, whether to commit them to your DMS, and how they interact with project memory.
- +Read the dedicated guide: /research/claude-md-files-guide
The Execution Layer
Dispatch, Recurring Tasks, and Sub-Agents
How Cowork actually runs work. These features govern when tasks start, how often they repeat, and how many agents work in parallel.
The Composition Layer
Projects, Skills, and Plugins
How you package and reuse work. The composition layer is where one-off tasks turn into repeatable systems.
Critical Distinction: Project vs. Standalone Session
- +This trips up almost every new Cowork user. The behavior is very different:
- +Standalone Cowork session: Pick a folder, start work. No memory, no custom instructions beyond Global, no scheduled tasks. Fresh slate every time.
- +Cowork Project: Dedicated workspace with persistent memory, project-specific instructions, attached folders, and scheduled tasks. Claude accumulates context across sessions.
- +Practical rule: If this is one-shot work, a standalone session is fine. If you will touch the same matter again next week — or want scheduled tasks tied to it — create a Project.
The Folder Selector at the Bottom
The folder picker at the bottom of a Cowork session controls what the VM can see. When you click it and choose a folder, Cowork mounts that directory into the virtual machine's filesystem. From that moment on, Claude can read, write, create, and delete files within that folder — but only within that folder. Files outside are invisible. If Claude needs to reach into a sibling directory, you either mount it too or move the relevant files into the current mount. Currently, the picker restricts selections to your home directory, which can be awkward for attorneys who keep client files on mapped drives or secondary volumes (this is a tracked limitation, not intentional).
Deep Dive Coming: Folder Best Practices
- +How you structure the folders you mount — one big bucket vs. per-matter directories, where to put CLAUDE.md files, what to exclude — has a large effect on Cowork's performance.
- +Read the dedicated guide: /research/claude-cowork-folder-best-practices
MCP: The Plumbing Layer
The Model Context Protocol (MCP) is Anthropic's open standard for connecting Claude to external tools and data sources. It is the plumbing beneath Plugins and Connectors — the protocol that lets Claude talk to iManage, Salesforce, a document database, or a custom internal system. If Skills are the instructions and Plugins are the apps, MCP is the wiring. You will interact with MCP most often when installing a connector (a remote MCP server) or when your firm's IT team stands up a local MCP server to expose an internal tool to Claude [7][8][11].
Deep Dive Coming: MCP Explained
- +MCP deserves its own walk-through — how to evaluate a connector, how to stand one up safely for a firm, and what the data-flow implications are for privileged matters.
- +Read the dedicated guide: /research/claude-mcp-explained
Cowork Is a One-Machine Tool
- +Cowork is designed for a single desktop — not a multi-device workflow. This shapes everything about how you plan to use it.
- +Memory lives on the machine: Project memory and scheduled tasks are stored on the specific computer running Cowork. Switching to your laptop means starting over there.
- +No hand-off: You cannot start a session on your desktop, pause, and pick it up on another machine. There is no shared state.
- +Mobile Dispatch is a partial exception: Pro and Max users can send new tasks from their phone to the desktop, but the desktop is still the one doing the work — it must stay open and awake.
- +Practical implication for firms: If attorneys switch between office and home computers, each needs its own Cowork setup. Plan for duplicated Global Instructions and separate Projects per machine, or standardize on one primary workstation.
Usage Limits
Cowork uses the same usage pool as Claude chat and Claude Code on your plan. Hitting the limit on one affects the others. The resets differ by tier: Free and Pro reset every 5–8 hours on a rolling basis; Max plans reset weekly. There is no unlimited tier — Max 20× raises the ceiling substantially but does not remove it.
Plan Tiers and Cowork Access
A snapshot of how Cowork availability maps to Claude plans as of April 2026. Prices are monthly. Verify at claude.com/pricing before purchasing.
Free
$0- +No Cowork access
- +Web and mobile Claude chat only
- +Useful for evaluation, not for Cowork workflows
Best for: Trial users deciding whether to subscribe
Pro
$20/month- +Cowork included
- +Dispatch access (available in beta)
- +Rolling 5–8 hour usage resets
- +Projects, Skills, Plugins, MCP
Best for: Solo practitioners and small-firm attorneys doing individual Cowork work
Max 5×
$100/month- +Cowork + Dispatch (Max was first to receive Dispatch at launch)
- +5× Pro usage allowance
- +Weekly resets
- +Extended Thinking
Best for: Attorneys and paralegals running Cowork for hours a day — document review, research, drafting
Max 20×
$200/month- +All Max 5× features
- +20× Pro usage allowance
- +Effectively all-day Cowork without hitting ceilings for most workflows
Best for: Power users treating Cowork as a continuous partner — litigation review, M&A diligence, multi-matter practices
Team
$25–$125/seat/month- +Cowork per seat (each user on their own desktop)
- +Premium seat unlocks Claude Code
- +Central billing and admin
- +No model training on firm data by default
Best for: Firms with 5+ attorneys standardizing on Claude for knowledge work
Enterprise
Custom- +All Team features
- +SSO, SCIM, audit logs, IP allowlisting
- +Custom data retention
- +HIPAA-ready offering available via BAA (BAA does not cover Cowork or Claude for Office — keep PHI workflows off those tools)
Best for: Firms with 50+ users needing compliance controls and custom terms
Current Limitations (from Anthropic's Help Center)
- +Some Cowork capabilities are not yet available:
- +Memory in projects only: Memory is supported within projects but is not retained across standalone Cowork sessions.
- +No chat or artifact sharing: Sessions cannot be shared with others.
- +Desktop app required: Cowork runs on your desktop computer via the Claude Desktop app. Pro and Max users can also message Claude from the mobile app while your desktop stays active. See Assign tasks to Claude from anywhere in Cowork for details.
- +Session persistence: The Claude Desktop app must remain open and your computer must be awake for Claude to work on tasks. If you close the app or your computer goes to sleep, active tasks will stop.
- +Source: Anthropic, 'Get started with Claude Cowork,' Claude Help Center.
Quick-Reference: Which Feature Do I Want?
A cheat-sheet within the cheat sheet. Match the problem to the tool.
I want Claude to behave a certain way in every session.
Global Instructions. Keep them under ~500 words.
I want rules that apply only to this client's folder.
Folder Instructions, or a CLAUDE.md file in that folder.
I want Claude to remember what we did last week.
Create a Project. Standalone sessions do not retain memory.
I want to send Claude a task from my phone.
Dispatch (Pro/Max). Desktop must stay open and awake.
I want a report every Monday morning.
Recurring Tasks (Scheduled Tasks) inside a Project.
I have a repeatable process I do identically each time.
Build a Skill. Works in chat, Cowork, and Claude Code.
I want a pre-built bundle for a practice area.
Install a Plugin — it wraps Skills, Connectors, and Sub-Agents.
I want Claude to read from my DMS or a firm system.
An MCP connector. Local for desktop-only tools, remote for internet-reachable systems (Team/Enterprise).
I want two parts of a task to run in parallel.
Sub-Agents — often set up for you inside a Plugin.
Getting Started Sensibly
A pragmatic sequence for a firm evaluating Cowork, minus the evangelism.
Install Claude Desktop and confirm your plan covers Cowork
Pro or above. Free does not include Cowork.
Write a minimal Global Instructions block
Your role, how you write, any absolute rules. Under 500 words. Resist the urge to list every preference — dilution is real.
Run your first standalone session against a low-stakes folder
Pick a non-privileged folder — a CLE outline, an internal memo, a research notes directory — and test a real task end to end.
Create your first Project when you hit the 'I wish it remembered' moment
Standalone sessions forget. The moment you notice yourself re-explaining context, promote that workstream to a Project.
Add a scheduled task only after the Project is stable
Recurring Tasks are high-leverage but only if the underlying task is reliable. Prove it manually first, then schedule.
Evaluate a Plugin before building Skills from scratch
Anthropic's plugin library covers common knowledge-work domains. Install before you author — the pre-wired Skills + Connectors will save hours.
Key Takeaways
- 1.Cowork runs inside an isolated Linux VM on your desktop — code never touches your host OS, and Claude can read/write only in folders you explicitly mount.
- 2.Three instruction layers stack: Global Instructions (every session) → Folder Instructions (per directory) → CLAUDE.md files (travel with the folder).
- 3.Three execution features govern how work runs: Dispatch (from phone), Recurring Tasks (on schedule), Sub-Agents (in parallel).
- 4.Three composition primitives turn ad-hoc work into systems: Projects (persistent memory), Skills (repeatable processes), Plugins (installable bundles).
- 5.Standalone sessions have no memory. If you want Claude to remember across conversations, you need a Project.
- 6.Cowork is single-machine. Memory and scheduled tasks live on the specific computer running the Claude Desktop app — no hand-off to another device.
- 7.The desktop app must stay open and the computer must remain awake for tasks to run. Sleep = stopped tasks.
- 8.Usage is shared across Claude chat, Claude Code, and Cowork on your plan. Free/Pro reset every 5–8 hours; Max resets weekly. See claude.com/pricing for authoritative limits.
- 9.HIPAA-ready Enterprise BAAs do not cover Cowork today. Firms processing PHI should keep that workflow outside of Cowork.
- 10.MCP is the plumbing beneath Plugins and Connectors — worth understanding before building firm-wide integrations.
References
- [1]Anthropic, "Get started with Claude Cowork," Claude Help Center.Link
- [2]Anthropic, "Organize your tasks with projects in Claude Cowork," Claude Help Center.Link
- [3]Anthropic, "Schedule recurring tasks in Claude Cowork," Claude Help Center.Link
- [4]Anthropic, "Assign tasks to Claude from anywhere in Cowork," Claude Help Center.Link
- [5]Anthropic, "Use plugins in Claude Cowork," Claude Help Center.Link
- [6]Anthropic, "What are Skills?" Claude Help Center.Link
- [7]Anthropic, "Getting Started with Local MCP Servers on Claude Desktop," Claude Help Center.Link
- [8]Anthropic, "Get started with custom connectors using remote MCP," Claude Help Center.Link
- [9]Anthropic, "Plans & Pricing — Claude by Anthropic."Link
- [10]Pvieito, "Inside Claude Cowork: How Anthropic Runs Claude Code in a Local VM on Your Mac," Jan. 2026.Link
- [11]Anthropic, "Introducing the Model Context Protocol."Link