MCP (Model Context Protocol) lets the model call external tools and data sources safely, such as fetching a page, looking up docs, or querying a database.
In CC GUI, open Settings → MCP servers. Full capability is centered on Claude Code.
Add a server
Click Add. There are three paths.
Manual config
Paste MCP Servers JSON. Prefer npx or uvx to start a local process.
Two transports:
Scope:
- Global: written to
~/.claude.json, available to all workspaces
- This workspace only: applies only to the current workspace
The config will run a command on your computer. Check the source, arguments, and environment variables before you add it.
Add from the MCP marketplace
The marketplace lists installable servers by registry. Search by name, publisher, or tag.
- Installable: can generate a config directly
- Browse only: metadata only. Add it yourself from the repo docs
Before you add one, check the command, URL, headers, and environment-variable placeholders. Unrecognized commands are marked as extra risk.
Import from GitHub Copilot
Paste Copilot MCP JSON. The root key is servers:
Preview first, then import. Duplicate names are renamed automatically and marked in the preview.
Built-in presets
Manual config also includes common presets:
- Fetch web content
- Time and time zones
- Knowledge graph / memory
- Sequential thinking
- Context7 docs lookup
A preset is only a starting point. Still confirm the command and permissions.
Each server card shows:
Open a card to:
- Enable / disable
- View the tool list
- Open connection logs
- Copy the config or the homepage / docs link
- Edit or delete
Refresh server status rechecks every server. The tool list uses cache first. Force a refresh when you need the latest data.
Use it in chat
After a server is enabled and connected, Claude Code picks tools from the task. You do not have to write the call in the input box.
Permissions still follow the current mode. In Default mode, an MCP tool may still ask first. See Modes and permissions. 
