Claude: Advanced

Advanced

At this level Claude stops being a chat box and becomes a hub that reaches your real tools and data. You’ll connect systems, extend Claude with MCP and Skills, and, critically for Swiss teams, get the data controls right.

Connect your data instead of pasting it

The connector directory gives you ready integrations: Google Drive, Gmail, Calendar, Slack, Notion, and more. A connected source keeps answers grounded and citable, and syncs across web, desktop, and mobile.

1. Open Customize > Connectors and pick a source from the directory.

2. Complete the sign-in / OAuth prompt to authorise it.

3. In a chat, ask Claude to pull from it directly:

   Find the latest version of the onboarding deck in Drive and list what changed since the March copy.

Grounding an answer in a live source beats pasting a stale export, and you get citations you can open and check. For a Project with a very large knowledge base, switch it to retrieval mode so Claude searches the files instead of trying to read all of them into the context window at once.

Extend Claude with MCP

MCP (the Model Context Protocol) lets Claude talk to tools the directory doesn’t cover. The desktop app gives you two paths:

Local (Desktop Extensions)

Single-click .mcpb packages at Settings > Extensions. Use Browse extensions for the directory, or Advanced settings to install a custom .mcpb. The desktop app ships with a built-in Node.js runtime, so most extensions need no separate install; Python and binary servers are supported too. Sensitive config such as API keys is encrypted in the macOS Keychain / Windows Credential Manager.

Remote (custom connectors)

For a hosted MCP server, go to Customize > Connectors > Add custom connector, enter the server URL, and complete OAuth if the server requires it.

This is context engineering made concrete: instead of describing your systems, you give Claude a tool to read them directly.

Build and use Skills

Skills are reusable instruction packages: a SKILL.md plus any supporting files. Claude invokes a Skill automatically when it fits, or you call it directly with a /slash command. You can upload one as a ZIP, and on Team/Enterprise share them org-wide.

Caution

Skills (and Artifacts file creation) require “Code execution and file creation” enabled at Settings > Capabilities. If it’s off, Skills silently don’t apply.

Custom Styles are being folded into Skills, so a house tone or output format becomes a Skill you and your team reuse.

Get your data controls right (Swiss teams, read this)

This is the trap for Swiss SMEs on individual accounts: consumer plans (Free, Pro, Max) use your chats for model training by default, a change that took effect on 28 August 2025. Leave the setting on and Anthropic retains your data for up to five years; turn it off and new chats fall back to the standard 30-day window. Commercial offerings sit under different terms entirely.

Plan	Trained on by default?	Retention	What to do
Free / Pro / Max	Yes (opt-out)	Up to 5 years if training is on; 30 days once you opt out	Turn off “Help improve Claude” at Settings > Privacy before any client data goes in
Team / Enterprise	No	Per commercial terms; a zero-data-retention addendum is available	Confirm your org’s admin settings; request ZDR where client contracts demand it
API (incl. Bedrock / Vertex)	No	Per commercial terms; ZDR available	Prefer it for anything programmatic that touches client data
Incognito chats (any plan)	Not used for training	Excluded from history and memory	Use for sensitive one-offs on a consumer plan

Your compliance path in three checks:

1. Individual account? Open Settings > Privacy and confirm “Help improve Claude” is off. The opt-out applies to new and resumed chats, not retroactively.
2. Company account? Verify the plan is Team or Enterprise and who controls the admin settings, and get the retention terms in writing.
3. Client-confidential work? Route it through a commercial plan or the API, or at minimum an opted-out account; use incognito for one-offs.

Caution

An individual Pro or Max account is not safe-by-default. Before putting client or confidential data into Claude, opt out of training or use a Team/Enterprise plan. This is a verification step for your whole setup, not just one answer. The table above matches Anthropic’s privacy documentation as of mid-2026; these policies move, so reconfirm before you commit to a client.

You’re done when…

• ✅ At least one connector is feeding Claude grounded, citable answers
• ✅ You’ve installed an MCP extension or a custom connector
• ✅ You’ve run or built a Skill, with code execution enabled
• ✅ Your training opt-out or plan choice is settled deliberately

Next step

Ready for the agentic edge with Cowork and computer-use?

Claude Expert Cowork and computer-use: real actions on your machine, scoped, supervised, and verified.