XNews

Anthropic Claude Code / Cowork and Agent Tooling

Anthropic Claude Code / Cowork and Agent Tooling

Anthropic launched Cowork, a GUI product extending Claude Code capabilities to non-programmers.

Key insights

Cowork grants folder-scoped file read/write access for automation

Cowork allows users to grant Claude access to a specific local folder; once authorized, Claude can read, edit, and create files within that folder. Typical use cases include converting screenshots into structured spreadsheets, organizing messy download folders, and drafting reports from scattered notes. Before major actions, Claude asks for user consent and reports progress while executing multi-step plans.

Skills, connectors, and shared ecosystem with Claude Code

Cowork can reuse Claude Code skills and user-defined Skills, and it integrates with existing connectors like Google Drive and other data sources. Skills are split into official and user-created types; Cowork loads skill metadata at startup and only fetches full content when matched, mirroring Claude Code’s skills architecture and enabling migration of existing workflows.

Cowork runs in a Linux VM causing some local script compatibility issues

Cowork executes inside a Linux virtual machine; on Apple M1/M2 ARM Macs this can break skills depending on local node_modules, browser cookies, or architecture-specific binaries. Pure-text skills and writing-style configurations work, but skills requiring local scripts or specific binaries may fail without additional adaptation.

Cowork supports multi-task queues and background execution

Users can queue multiple tasks for Claude to execute in the background; Claude will autonomously plan, execute steps, and provide progress updates. This model enables parallel task processing without requiring the user to remain in an active chat session.

Security model limits access but requires careful instruction phrasing

Claude only accesses explicitly authorized folders and connectors and prompts before major operations, but user reports note risks: ambiguous instructions can cause unintended deletes, and prompt-injection via external content remains a concern. Anthropic has protections but advises cautious initial use and permission-limited skills.

Source posts