Anthropic debuts Cowork for Claude on macOS

Overview

Anthropic has released Cowork, a research preview in the Claude macOS app that brings the “Claude Code style” of agentic work to everyday documents and files. Instead of working only inside a chat window, Cowork lets you grant Claude access to a specific folder on your Mac so it can read, edit, and create files directly in that workspace.

Cowork is available now for Claude Max subscribers in the macOS app, with Anthropic signaling rapid iteration based on early user feedback. The company also says it plans to add cross device sync and expand to Windows in future updates.

What Cowork is (and how it differs from chat)

In a regular conversation, you paste context in and manually move outputs into the right format. In Cowork, Claude works more like an assistant operating on real artifacts:

• You select a folder Claude can access
• Claude can create and modify files in that folder
• Claude can form a plan, execute step by step, and keep you updated
• You can queue tasks so Claude can work through them in parallel

Anthropic positions Cowork as “Claude Code for the rest of your work,” built on the same agentic foundations used by Claude Code.

Key features in the research preview

Cowork focuses on practical file based automation and document generation. Notable capabilities include:

• Folder level access controls: Claude can only see folders you explicitly allow
• Real file operations: read, edit, create, rename, reorganize, and potentially delete files
• Agentic execution: Claude makes a plan and completes tasks with less back and forth
• Connectors support: Cowork can use existing Claude connectors to pull in external information
• Skills for work outputs: an initial set of Agent Skills aimed at improving creation of documents, presentations, and other files
• Optional browser pairing: with Claude in Chrome, tasks can extend to browser based steps

Example tasks Anthropic highlights include reorganizing a downloads folder, extracting expenses from screenshots into a spreadsheet, or drafting a report from scattered notes.

Stay in control: safety and risks to understand

Anthropic emphasizes user control, but also acknowledges new risks that come with letting an AI agent touch real files.

• Explicit permissions: Claude cannot access anything outside folders and connectors you approve
• Confirmation prompts: Cowork asks before significant actions, giving you a chance to steer
• Destructive actions are possible: if instructed, Claude could delete local files, so clear guidance and guardrails matter
• Prompt injection remains a concern: content from the web can try to redirect an agent’s behavior; Anthropic says it has defenses, but agent safety is still an active industry problem

For teams evaluating Cowork, it can help to view this release alongside Anthropic’s broader work on agent behavior and reliability, including Anthropic’s recent release of the Bloom framework, which focuses on evaluating and stress testing model behavior.

For teams adopting Cowork, the practical takeaway is to treat it like any automation tool: start in a low risk folder, use backups or versioning where possible, and define rules for what the agent can change.

Impact for developers, founders, and teams

Even though Cowork targets non developers, the release is a meaningful signal for builders:

• UI pattern shift: agentic work is moving from chat to “workspace plus tools,” where the model operates on real assets. This mirrors the broader shift toward agent-to-UI workflows showing up across the industry.
• Workflow prototyping: founders can test how far agent behavior can go in document heavy ops without building custom tooling
• Connector plus Skills stack: developers can view Cowork as a reference implementation of how connectors and agent Skills combine into repeatable workflows

Availability and what’s next

Cowork is live as a research preview for Claude Max users via the Claude macOS app (look for “Cowork” in the sidebar). Anthropic is collecting feedback to guide improvements, including planned Windows support and cross device sync.

As with any early agent feature, the most useful approach is to start small, validate results, and gradually expand the scope of what you let the agent handle.

Discover more cutting edge AI tools and apps on Appse, your go to directory for the latest AI innovations.

Source: Anthropic: Cowork research preview