Claude Cowork is a very restricted version of Claude Code

Claude Cowork is not simply "Claude Code with a nice UI". It's a sandboxed version of Claude Code with major restrictions, that aim to prevent non-technical users from getting into trouble. Many "power-user" workflows cannot be implemented in Cowork.

Below are some of the limitations I've found so far.

Network access

WebFetch can get text from a URL, but not HTML.

It can make GET requests only—no POST, no custom headers.

This means you can only interact with services that have an official connector, or an MCP you can setup manually. You can't run skills that call external APIs, e.g. my invoicing skill that queries my time tracking software and creates invoices in Xero, or my ask many models skill that queries models from Google, OpenAI, Grok and Anthropic all at once.

No OAuth or CLI auth flows

CLI tools that require browser-based authentication won't work in the sandboxed VM. This includes:

• gh auth login (GitHub CLI)
• gcloud auth login (Google Cloud)
• aws sso login
• npx vercel login

If your skill depends on authenticated CLI tools, it won't run in Cowork.

Limited package managers and tools

Browser automation has heavy guardrails

Claude in Chrome exists in Cowork, but with significant restrictions. Many actions require explicit per-action confirmation in chat—even if you've already described your intent.

This makes Cowork unsuitable for fully autonomous browser workflows. Every "irreversible" action becomes a back-and-forth.