Cowork is a new research preview feature for Claude Max subscribers, available on the macOS app, designed to extend Claude's capabilities beyond conversational AI into more autonomous, agentic work with local computer files. Developed from the observation that developers were already using Claude Code for diverse tasks beyond coding, Cowork aims to provide similar functionality in a more approachable form for non-developers and non-coding tasks.

The core methodology of Cowork centers on granting Claude controlled access to a user-specified local folder. Within this designated environment, Claude is empowered to read, edit, and create files. This capability enables Claude to perform complex, multi-step tasks with significant agency, such as:

1. File Organization: Analyzing and re-organizing local files, e.g., sorting and renaming downloads.
2. Data Extraction and Synthesis: Processing unstructured data from various formats (e.g., extracting expense lists from a pile of screenshots) and generating structured outputs (e.g., a spreadsheet).
3. Content Generation: Producing first drafts of documents (e.g., a report from scattered notes).

Technically, Cowork builds upon the foundational architecture of Claude Code, which facilitates agentic behavior and tool use. The system functions by allowing Claude to:

• Plan and Execute: Upon receiving a task, Claude formulates an internal plan and systematically works through it, providing updates to the user. This represents a significant departure from typical back-and-forth conversational models, shifting towards an asynchronous, task-oriented workflow.
• Leverage Connectors: Cowork integrates with existing "connectors," which are mechanisms that link Claude to external information sources, enabling it to access and utilize data beyond the local environment.
• Utilize Skills: An initial set of "skills" has been added to enhance Claude's ability to create specific document types, presentations, and other file formats. These skills likely encapsulate specific prompts, tool-use sequences, or fine-tuned models for structured output generation.
• Browser Access: When paired with "Claude in Chrome," Cowork can extend its agency to tasks requiring browser interaction, such as data gathering from websites.
• Asynchronous Task Management: Users can queue multiple tasks for Claude, which it can process in parallel, reducing waiting times and enabling a more continuous workflow akin to delegating tasks to a human coworker.

Safety and control are emphasized through explicit user permissions. Claude can only access folders and connectors that the user explicitly grants access to, ensuring no unauthorized data access. Furthermore, Claude is designed to ask for confirmation before executing significant or potentially destructive actions, such as deleting local files. Users are advised to provide clear instructions to mitigate the risk of misinterpretation by the AI. The paper also acknowledges the ongoing industry challenge of "prompt injections"—attempts to maliciously alter an AI's behavior through manipulated input—and states that sophisticated defenses are in place, but agent safety remains an active research area.

Cowork is released as a research preview to gather user feedback and iterate on its capabilities. Future plans include improvements such as cross-device synchronization, support for Windows operating systems, and further enhancements to safety mechanisms.