# Claude Cowork Claude Cowork is [[Anthropic]]'s general-purpose agent tool, launched in January 2026. Extends [[Claude Code]]'s agentic capabilities to non-developers. Accessible as a dedicated tab in Claude Desktop alongside Chat and Code. Described as "Claude Code for the rest of your work." ## Key features - File access: read, edit, create, and organize files within user-approved folders - Web searching via built-in search tool - Terminal command execution and code running - Artifact generation (interactive HTML, visualizations) - Multi-step task execution with progress tracking - Task queuing and parallel processing - Compatible with Claude connectors and [[Claude in Chrome]] ## How it works Claude develops and executes plans autonomously while keeping users informed. Unlike standard chat, users don't need to manually provide context or convert outputs. Tasks run locally in a sandboxed environment. Under the hood, uses Apple's VZVirtualMachine framework running a custom Linux root filesystem. ## Example use cases - Reorganizing and renaming files in downloads - Creating spreadsheets from scattered screenshots - Drafting reports from unorganized notes - Reviewing and analyzing document collections - Web research with multi-step verification ## Safety and control - Filesystem sandbox: Claude can only access user-approved folders - Permission prompts before significant actions - Containerized file access with default sandbox isolation - Summarization layers to reduce prompt injection risk - Agent safety remains an active area of development ## Dispatch (mobile-to-desktop task assignment) Dispatch gives a single persistent conversation with Claude that syncs across phone and desktop. Assign Claude a task from mobile, go do something else, and come back to the finished work. Claude runs on the desktop computer with access to local files, connectors, and plugins, and messages the result when done. ### How it works - One continuous thread that doesn't reset; Claude retains context from previous tasks - Message from phone on the way to work, follow up from desktop when you sit down - Claude works on the desktop using files, connectors, and plugins already configured in Cowork - Claude messages the outcome (spreadsheet, memo, comparison table) rather than showing every step ### Requirements - Claude Desktop app (macOS or Windows x64) installed and running (computer must be awake) - Claude mobile app (iOS/Android), latest version - Pro or Max plan - Active internet connection on both devices ### Setup 1. Open Cowork on phone or desktop 2. Click "Dispatch" in the left side panel 3. Follow the setup flow; optionally grant file access and keep-awake ### Example use cases - Pull data from a local spreadsheet and compile a summary report - Search Slack messages and email, then draft a briefing document - Build a formatted presentation from Google Drive files - Organize or process files in a specific folder on the computer ### Security considerations Mobile agent remotely controls a desktop agent. Instructions from the phone can trigger real actions on the computer: reading, moving, or deleting local files, interacting with connected services, controlling the browser. Manipulated instructions or phishing content encountered along the way could cascade into irreversible actions. ### Current limitations (research preview) - Desktop must be active (computer awake, app open) - Claude responds only to messages; no proactive outreach - Single continuous thread only; no multiple threads - No notifications when tasks complete - No scheduled tasks within the Dispatch thread (managed separately) ## Record a skill (July 2026) The most interesting addition since launch, and the one worth thinking about properly. **You record your screen while doing a task, talking through it as you go, and Claude turns the recording into a skill it can run again.** It sits under *Record a skill* in the `+` menu of the desktop app. Pro, Max and Team plans. ### Why this is a bigger deal than it looks Every skill I write today is written the hard way: I do the task, notice I'll repeat it, then sit down and describe it in text, in the right structure, with the right triggers. That authoring step is the tax, and it's why most repeated work never becomes a skill at all. The work gets done a dozen times and codified zero. Recording collapses that. **Demonstration replaces specification.** You do the thing once, narrating, and the artifact falls out. The narration is doing the real work. A screen recording alone gives you actions without intent, which is exactly the wrong half: it's the *why did I click that* that makes a skill generalise instead of replaying one brittle path. Talking through it supplies the reasoning the pixels can't. ### The obvious question How well does it generalise? A recording is one path through one instance of a task, with one set of files and one set of edge cases. The gap between "replay what I saw" and "handle the case I didn't demonstrate" is the whole game, and it's the same gap that makes [[AI Agent Skills]] hard to write well by hand. Worth testing directly: record a task, then run it against deliberately different input and see what survives. That single experiment tells you whether this is a genuine authoring shortcut or a macro recorder with better marketing. ### Why it matters to me I maintain a large skill library, all hand-written. If recorded skills produce a usable first draft that I then edit, the authoring cost of a new skill drops sharply and the set of things worth codifying gets much larger. Marginal skills that were never worth twenty minutes of writing become worth three minutes of narrating. It also points at where this is heading generally. First we wrote programs, then we wrote prompts, now we demonstrate and let the machine write the prompt. Each step moves the specification further from code and closer to how you'd explain a task to a colleague. ## Availability - macOS and Windows x64 - Pro ($20/month) and Max ($100-$200/month) plans; Team/Enterprise plans - Research preview status - *Record a skill*: Pro, Max and Team ## References - Announcement: https://claude.com/blog/cowork-research-preview - Record a skill announcement (2026-07-21) — https://x.com/claudeai/status/2079595988998554047 - Simon Willison's analysis: https://simonwillison.net/2026/Jan/12/claude-cowork - Dispatch (mobile task assignment): https://support.claude.com/en/articles/13947068-assign-tasks-to-claude-from-anywhere-in-cowork - Safe usage: https://support.claude.com/en/articles/13364135-use-cowork-safely - Scheduled tasks: https://support.claude.com/en/articles/13854387-schedule-recurring-tasks-in-cowork ## Related - [[Claude Code]] - [[Claude in Chrome]] - [[Anthropic]] - [[AI Agents]] - [[AI Agent Skills]] - [[Claude Code Skills]] - [[AI Assistants]] - [[OpenWorker]] - [[Claude Desktop on Linux]]