Cowork

Best for: Adding capability without writing code — drop in a skill, it triggers when relevant.

• Browse and install from the in-app Skills Directory
• Each skill has a description (trigger) and a body (playbook)
• Skills can ship supporting files — templates, example outputs, scripts
• Build your own with the skill-creator skill (built in)
• User skills coexist with plugin-shipped skills and Anthropic example skills

Limitations: The directory grows fast — discovery is a real problem. If two skills could fire on the same prompt, the more specific description wins; vague descriptions misfire. Read SKILL.md before installing third-party skills you'll let auto-fire.

Best for: Onboarding a function (sales, finance, legal, marketing) in one click.

• 11 starter plugins from Anthropic, plus a growing community marketplace
• Each plugin is a one-click install (or one CLI command for advanced users)
• Plugins can be customized per organization — the cowork-plugin-customizer skill walks through it
• Plugin Create tool generates a working plugin from a short brief
• Versioned and revertible — pin a known-good build for a team

Limitations: Plugins inherit the trust of everything they ship — skills auto-fire, MCP servers reach external systems. Audit before installing a third-party plugin, especially ones with connectors that write to your data.

Best for: Letting Cowork read and act on your real work data — calendar, email, CRM, docs.

• MCP registry built into Cowork — search by app name or capability
• OAuth authentication flow handled in-app per connector
• Connectors expose typed tools (e.g., gmail.list_drafts, calendar.create_event)
• Same protocol as Claude Code MCP — server portability across surfaces
• Suggest-connectors hint surfaces relevant connectors when a task mentions an app

Limitations: Connector tools run with the agent's authority — a compromised or misbehaving connector is a real risk vector. Anthropic-listed connectors are reviewed; third-party servers should be treated like dependencies. Read-only scopes where possible.

Best for: Turning a one-off output into a recurring routine.

• Trigger with /schedule or natural language ("every morning at 6am, give me a news briefing")
• Cron-style schedules or one-off fireAt times
• Output lands as a normal conversation turn — no separate inbox
• List/update/delete via /schedule list and update_scheduled_task
• Common patterns: morning briefing, weekly digest, hourly inbox sweep

Limitations: Desktop-bound: tasks only run when the app is open and the machine is awake. Anthropic's Routines (separate research-preview offering) is the cloud-hosted answer for runs that need to happen with the laptop closed.

Best for: Turning a one-off answer into a page you keep coming back to.

• create_artifact saves an HTML page that survives session restarts
• window.cowork.callMcpTool(name, args) calls any connector tool
• window.cowork.askClaude(prompt, data) runs lightweight inference on demand
• CDN-allowed libraries: Chart.js, Grid.js, Mermaid (other JS must be inline)
• Best use cases: status pages, trackers, recurring reports, interactive explorers

Limitations: Artifacts re-fetch data — they're not snapshots. If the underlying connector goes down, the artifact shows the error. Build for graceful failure. Don't store secrets in artifact HTML; it's stored as-is.

Best for: File-heavy work where the agent needs to touch real files, not paste content.

• Per-folder approval — you see and approve the path
• Defaults can be set in user preferences (auto-mount each chat)
• Read, Write, Edit, Grep, Glob tools all work inside an approved folder
• Bash runs in an isolated Linux sandbox with the folder mounted at /sessions/...
• Outputs folder is the scratch space; mounted folders are the real workspace

Limitations: The agent has full read/write inside an approved folder — there is no sub-path access control. If you mount your whole home directory, the agent can read every file in it. Mount narrowly.

Best for: Web automation: configure SaaS dashboards, fill forms, copy credentials, run JavaScript-rendered pages.

• Tools: navigate, get_page_text, read_page, form_input, file_upload, javascript_tool
• Reads what the user sees (rendered DOM) — works on client-side React/Vue apps
• Can run JavaScript in the page for advanced inspection
• Right escalation when WebFetch returns a page shell or "enable JavaScript" message
• Good for: Supabase setup, Vercel config, copying API keys, filling job-application forms

Limitations: Slower than an MCP connector when one exists — if the app has a connector, use it. Browser sessions tie up a real browser instance. Treat browser-driven actions like manual clicks: write-with-care.