Claude Cowork capabilities

Extract, organize, and format action items from meeting transcripts and documents.

Autonomous multi-step workflow execution on files, documents, and data through the Claude desktop app.

Cowork 'live artifacts' are persistent, interactive HTML pages (trackers, dashboards, comparison tools, morning briefs) that survive across sessions and can call the user's MCP connectors at open time to repaint with fresh data. They are distinct from claude.ai chat Artifacts, which are single-session and don't call MCPs live.

Claude Cowork is Anthropic's desktop agent for non-developer knowledge workers: it brings Claude Code's agentic, multi-step task execution into the Claude Desktop app so people in sales, finance, legal, marketing, ops, HR, and design can hand off complex work that touches local files, web data, and connected SaaS tools.

Cowork operates on user-selected local folders that the user explicitly grants Claude read/write/edit access to. Claude can read every file inside, write new files, and modify existing ones, producing real .docx/.xlsx/.pptx/.pdf outputs in place rather than chat-only artifacts.

The primary surface is the Claude Desktop app, which has a mode selector toggling between 'Chat' and 'Cowork' (Tasks). Cowork tasks render inline in the desktop window with a sidebar for tasks, scheduled jobs, live artifacts, and the Customize menu.

Read, write, and organize files on the user's local file system through the Claude desktop app.

External SaaS access in Cowork is delivered through MCP (Model Context Protocol) connectors — remote MCP servers the user authorizes via OAuth that let Claude read and act in tools like Google Drive, Gmail, Slack, Asana, Notion, Supabase, Vercel, Airtable, and more.

Generate structured meeting summaries from uploaded transcripts, recordings, or raw notes.

Cowork uses Anthropic's frontier models with a shared picker — the same model lineup users see in regular Claude chat (as of 2026-06-12 claude.com/pricing lists Mythos / Fable / Opus / Sonnet / Haiku, top to bottom; Mythos is the newest top tier and not yet publicly characterized), with Sonnet as the default workhorse and the larger models reserved for harder reasoning.

Analyze and cross-reference information across multiple documents, files, and data sources in a single task.

First-run flow for Cowork that asks the user what role they have (sales, finance, marketing, ops, etc.) and uses that to recommend a starter set of plugins, connectors, and Skills tailored to that function — the bridge between 'fresh install' and 'useful agent.'

Cowork is gated behind paid Claude plans only — Pro ($20/mo, $17/mo annual), Max 5x ($100/mo), Max 20x ($200/mo), Team, and Enterprise. Free Claude users do not get Cowork. Originally Max-only at January 2026 launch, Pro access was added shortly after with lower usage limits than Max.

Plugins are role-shaped bundles that package Skills, MCP connectors, and sub-agents together so a user gets a working setup from the first conversation instead of wiring pieces individually. Anthropic ships a library covering sales, finance, legal, marketing, HR, engineering, design, operations, data analysis, and more.

Projects in Cowork are persistent, self-contained workspaces that group related tasks, files, links, instructions, and memory together — the unit of organization for recurring or long-running work like 'Q3 board prep' or 'monthly close.'

Shell commands and any code Claude writes execute inside a dedicated, isolated Linux VM, separate from the host OS, with network egress filtering, syscall restrictions, and per-session user isolation. The native agent loop (conversation logic, connected-folder file operations, web requests, local plugin MCP servers) runs on the host outside the VM.

Scheduled tasks let a user describe a recurring job once — e.g. 'every Monday at 9am, build a weekly metrics report from these connectors' — and have Cowork run it automatically on that cadence without re-prompting.

Skills are folders of instructions, scripts, and resources (SKILL.md plus optional assets) that give Claude a repeatable workflow or specialized knowledge for a task, loaded automatically when the model decides it's relevant. They appear in Cowork via plugins or as individually installed Skills from the unified directory.

Slash commands are user-triggered shortcuts in the Cowork composer that invoke a structured workflow or built-in action. Typing / shows every built-in and installed skill the user can invoke explicitly — current docs treat / as the explicit way to launch Skills, complementing their automatic loading, rather than a fully separate system.

Cowork can split work across parallel sub-agents via a Task tool, letting the main agent fan out independent subtasks (research, file processing, parallel writes) and aggregate results, turning serial work like 'analyze 10 files' into concurrent work.