What Changed — Direct OS Control via Claude Code
Anthropic has released a "computer use" feature for Claude, enabling it to directly control your Mac's operating system. For Claude Code users, this functionality is accessible through a new /control flag in the terminal. This isn't a separate app—it's integrated directly into the Claude Code workflow you already use for coding tasks.
The feature works by granting Claude temporary, permission-based access to perform actions like opening applications, navigating files, clicking UI elements, and executing system commands. It operates within a sandboxed environment with explicit user approval required for sensitive operations.
What It Means For Your Development Workflow
This fundamentally changes how you can structure your coding sessions. Instead of context-switching between your terminal and other apps, you can now delegate those tasks to Claude while you focus on core development.
Concrete examples for developers:
- While debugging a Node.js server, have Claude open Chrome, navigate to
localhost:3000, and click through your UI to reproduce an issue - During a complex refactor, tell Claude to "open the project in VS Code, find all imports of
lodash, and show me the files in Finder" - When setting up a new project, delegate environment setup: "Check if Docker is running, launch PostgreSQL from Applications, and verify the port"
This follows Anthropic's March 2026 release of Claude Code Auto Mode and the /dream command for memory consolidation, continuing their push toward more autonomous developer tools.
How To Use It Safely Right Now
Installation & Setup:
# Update to the latest Claude Code version
claude code update
# Enable computer use with explicit permissions
claude code --enable-control --permissions-file ./claude-perms.json
Create a claude-perms.json file to limit access:
{
"allowed_apps": ["Finder", "Chrome", "VS Code", "iTerm"],
"allowed_folders": ["/Users/yourname/Projects", "/Users/yourname/Downloads"],
"require_approval": ["system_preferences", "terminal_admin_commands"]
}
Sample workflow in action:
# Start a coding session with control enabled
claude code --control
# Then in the Claude Code interface:
> I need to test the login flow. Can you open Chrome to localhost:8080,
> click the login button, and tell me what error appears?
# Claude will request permission for each action, then execute them
# while you continue working in another terminal tab
Critical safety practices:
- Always use the permissions file—never run with unrestricted access
- Start with specific folders—don't grant access to your entire home directory
- Monitor the approval prompts—Claude will ask before each new type of action
- Use separate project directories—keep sensitive files outside controlled folders
When This Beats Your Current Tools
This isn't about replacing your entire workflow—it's about strategic delegation:
Use /control when:
- You're in deep focus mode and don't want to break concentration
- You need repetitive UI testing across multiple browsers
- You're working with unfamiliar system configurations
- You want to document exact reproduction steps for bugs
Stick to manual control when:
- Working with sensitive production data
- Performing irreversible system operations
- Speed is critical (manual is still faster for simple tasks)
The Bigger Picture: Agentic Development Environments
This release represents another step toward what we've been calling "agentic development environments"—where Claude Code handles not just your coding tasks but the entire surrounding workflow. As noted in our March 27 coverage of Satya Nadella's predictions about AI agents commoditizing traditional SaaS, the value is shifting to the orchestration layer. Claude Code with /control becomes that orchestration layer for your local development environment.
What makes this particularly powerful for developers is that it builds on the existing Claude Code infrastructure you're already using. The MCP (Model Context Protocol) servers you've configured for coding tasks can now be coordinated with OS-level actions, creating truly end-to-end automation workflows.
