Cowork Glossary

Describes AI systems that autonomously plan and execute multi-step tasks with minimal human intervention. Cowork is "agentic" because it analyzes your request, creates a plan, and executes it with your approval.

Research preview feature enabling multi-agent coordination where multiple Claude agents work together on complex tasks. One agent orchestrates others, each handling a specialized sub-task in parallel.

The checkpoint where Cowork shows its planned actions and waits for your confirmation before executing. Critical security measure — always review before approving.

A file created by Claude (document, spreadsheet, image, etc.) during a Cowork session. Artifacts appear in Claude's interface where you can preview, edit, or download them.

The macOS and Windows application that hosts Cowork. Different from the web interface (claude.ai).

Beta feature that automatically compresses conversation history during long sessions. Allows much longer effective sessions by keeping relevant context while removing redundant information.

The maximum amount of text/data Claude can process in a single session (~200K tokens). When exceeded, tasks may fail or produce incomplete results.

The "memory" space where Claude holds your conversation and file contents during a session. Measured in tokens.

Claude's agentic desktop feature for knowledge workers. Accesses and manipulates files directly on your computer without requiring code execution.

The settings area in Claude Desktop for managing Skills, Connectors, and personalizations. Install skills, configure external tool connections, and set per-tool permissions.

Official Cowork extension installed from the Customize tab. Enables file system access beyond the workspace folder, cross-session memory via a memory.md file, and one-click MCP server installation.

The detailed list of actions Cowork proposes before starting work. Shows what files will be affected and how. Review this carefully before approving.

Claude's ability to reason through complex problems step-by-step before responding. Produces more accurate results for multi-step tasks at the cost of slightly longer wait times.

The restricted area where Cowork can operate. You grant access to specific folders; Cowork cannot access anything outside this boundary. Core security mechanism.

Non-technical professionals who work primarily with information and documents — project managers, analysts, writers, consultants. Cowork's primary target audience.

Design principle where data and processing stay on your computer rather than in the cloud. Cowork accesses local files only — nothing is uploaded without your explicit action.

Model Context Protocol integration allowing Cowork to interact with external tools and services. Three types: web search, desktop/local files, and custom JSON. No coding required for setup.

A markdown file (typically memory.md) used with Desktop Commander to persist context across sessions. Contains business context, client preferences, and recurring task information.

An operation requiring multiple sequential actions, e.g. "read files → analyze → create report → organize output." Cowork excels at automating these without requiring code.

Optical Character Recognition. Technology that extracts text from images and scanned documents. Cowork uses OCR to read receipts, screenshots, and PDFs that are image-based.

The main Cowork agent that receives your request, creates the execution plan, coordinates sub-agents, and assembles the final result.

Official third-party integration that extends Cowork's capabilities. Examples: Asana, Canva, Figma, GitHub, Google Drive, Jira, Notion, Slack. Allows direct service interaction without browser automation.

The instructions you give to Claude. Can range from simple ("organize these files alphabetically") to complex multi-step workflows. Clear, specific prompts produce better results.

A security attack where malicious instructions are hidden inside files (PDFs, documents) attempting to manipulate Claude's behavior. Mitigation: only process files from trusted sources.

Feature allowing automated recurring Cowork operations on a set schedule — e.g. daily report compilation, weekly file organization. Removes the need to manually trigger repetitive tasks.

Add-on capability for Cowork, installed via the Customize tab. Invoked via slash commands (e.g. /pdf, /docx, /xlsx). Skills extend what Cowork can do beyond its defaults.

Combining multiple skills in a single workflow. Example: /pdf to extract content from a document, then /xlsx to organize it into a spreadsheet. Skills execute sequentially.

A shortcut for invoking a skill in Cowork. Type the slash command at the start of your message — e.g. /pdf, /docx, /xlsx, /canvas-design.

Specialized worker spawned by the orchestrator to handle a specific part of a task. Each sub-agent has fresh context and can work in parallel with others.

The unit Claude uses to measure text. Roughly 4 characters or 0.75 words. Used to calculate context limits and usage costs.

A repeatable sequence of actions that transforms input into output — like a recipe you can reuse. Example: "read files → process → create output." Cowork can automate 28+ documented workflows.

The dedicated folder structure for Cowork operations. Best practice: ~/Cowork-Workspace/ with input/ and output/ subfolders. Cowork operates within this boundary by default.