Anthropic Releases Cowork as Claude’s Local File System Agent for Daily Work

Anthropic released Cowork, a new feature that uses agent workflows in local files for uncoded tasks available in a research preview within the Claude macOS desktop application.

What a Workaround Does at the File System Level

Cowork currently works as a dedicated mode for the Claude desktop app. When you start a Cowork session, you select a folder on your system. Claude can only read, edit, or create files in that folder.

Anthropic provides concrete examples. Claude can rearrange the downloads folder by sorting and renaming files. It can read a directory of screenshots, extract values, and create a cost spreadsheet. It can cut through the scattered notes in that folder and generate a ready-made report.

The interface keeps communication in a common conversational environment. You describe the function in natural language. Claude builds the internal system, performs file operations, and broadcasts status messages as it progresses. You can continue to send follow-up instructions while the job is in progress.

Relationship With Claude Code And Claude SDK Agent

Cowork is not a unique model. Anthropic says Cowork is built on the same foundations as Claude Code and uses Claude’s agent SDK as the underlying agent stack.

Claude Code began as a command-line oriented environment that allowed developers to run shell commands and modify project files using natural language, with the latest web interfaces and Slack on top. Many Max users have pushed Code Claude into non-coding workflows, using it as a general-purpose agent that runs on directories and tools arbitrarily. That usage pattern directly informs Cowork.

TechCrunch describes Cowork as a more accessible version of Claude Code, used as a sandboxed example of a co-agent stack. Users still select a specific folder, but there is no need to work on storage or configure physical locations.

Browser-Based Connectors, Capabilities, and Flows

Collaboration can go beyond local storage. Anthropic allows Cowork to reuse Claude’s existing connectors, which include external services such as Asana, Notion, and PayPal. Cowork also supports the first set of Skills developed for creating documents, presentations, and similar artifacts. These instructions package skills and resources for specific job functions.

When Cowork is paired with Claude in Chrome, the same agent program can include browser steps. Articles by Anthropic and The Verge both highlight browser-related functions where Claude can follow links, read pages, and act within web applications under user supervision.

This combination gives Cowork a three-layer toolkit:

1. Local file system access is limited to the selected folder.
2. Connectors for external systems are in order.
3. Browser actions by Claude in Chrome.

From an implementation perspective, this is a standard agent tool configuration on top of the Claude agent SDK. The difference is that Cowork hides the tool graph and only exposes the chat task interface.

Agent Behavior, Scheduling, and Related Functions

Anthropic emphasizes that Cowork is working with more agencies than Claude’s usual interview. Once you’ve specified a task, Claude creates a plan, executes a series of tool calls and file operations, and keeps you up-to-date on intermediate steps.

You don’t need to paste content over and over again or manually modify the output. A coworker reuses a folder as a persistent context boundary. It writes intermediate artifacts directly to that directory, and uses those artifacts in subsequent steps.

A Model for Security, Access Control, and Fast Injection

Because Cowork works on real files, security constraints are clear in product design. Anthropic says users choose which folders and connectors Claude can see. Claude cannot read or edit content without those properties.

A colleague always asks for confirmation before taking important steps. Yes, with all the benefits, there are heads-up: It is recommended to give accurate instructions and warn that incorrect interpretations are possible.

Anthropic also cites rapid injection as a primary risk. If Claude processes untrusted content on the Internet or in local documents, that content can try to change the system and avoid behavior away from the user’s intent. Anthropic claims to have implemented safeguards, but describes agent security, which is defined as protecting real-world actions, as an ongoing research problem.

Availability

Cowork is available today as a trial preview for Claude Max subscribers using the macOS desktop system. Max’s plan costs between $100 and $200 dollars per month depending on usage. Users on other programs can join the waiting list. Anthropic plans to add device sync and Windows support in future iterations.

Key Takeaways

• Local folder scope agent: Cowork runs within the Claude macOS operating system as an agent that can read, edit, and create files only within user-selected folders, giving Claude direct file system access through a clear program.
• Same stack as Claude code, different location: A unified task is built on top of the Claude agent SDK basics like Claude Code, but it directs non-coding workflows through a GUI instead of a backend-based developer interface.
• Tools: files, connectors, browser: Cowork combines three tool layers in one program, local file functionality, Claude connectors for services like Notion or Asana, and browser actions with Claude in Chrome.
• Agentic, multi-step execution: Once a task has been assigned, Cowork organizes and executes a multi-step workflow, broadcasts progress updates, and can queue multiple tasks in parallel, instead of acting as a single rapid response loop.
• It is forced but has the power to destroy: Access is limited to selected folders and configured connectors, and Cowork asks before major actions, but it can still perform destructive operations such as file deletion, so it should be considered a powerful automation tool, not a harmless chat bot.