Extensions (MCP)

Give the agent new tools with extensions — built-ins work out of the box, third-party MCP servers install only with your explicit consent.

What are extensions

Extensions are plugins based on the MCP (Model Context Protocol): they connect external apps and tools to the agent so the model can read data, operate your system, or call services — file systems, browsers, databases, memory, and more.

Built-in extensions (out of the box)

The app ships with a set of built-in extensions you can toggle individually in "Settings → Extensions":

  • Developer (enabled by default): general development tools — reading and writing files, running commands, and more;
  • Memory: lets the agent remember your preferences as you work;
  • Plus other assistive capabilities (tutorial, visualisation, etc.).

Adding third-party MCP servers

Any standard MCP server can be added as an extension (a stdio command or an HTTP endpoint):

  1. In "Settings → Extensions" choose Add, then enter the name and command / URL.
  1. Installation requires your explicit consent — before confirming, the app shows the extension's source, type, and scope of capabilities.
  1. External extensions are automatically checked for known malware packages before activation; a hit is blocked with a clear error.

Safety note: extensions are programs that run on your machine — only install extensions from sources you trust; after installation their tool calls are still governed by the Allow / Ask / Deny permission modes (see Permissions & safety).

The session-level browser extension

The embedded browser in the right-hand workspace appears to the model as a "browser" extension: it is off by default, and only after you enable the master switch in "Settings → Chat → Browser control" and authorize domains one by one can the model read or act on web pages. The extensions page shows that entry as status-only; the switch lives in chat settings.

Next steps

  • Want the agent to follow domain know-how? See Skills.
  • Want to combine extensions into reusable tasks? See Workflows.

This page is adapted from upstream open-source documentation (Apache-2.0) with modifications; see /legal/provenance for sources and changes.