Skip to content

Skills, MCP and plugins

The extensions panel gathers four tabs: Skills, MCP, Plugins and LSP. They all serve the same idea: giving the agent new capabilities without you writing any code.

A Skill is a package of reusable instructions (“how to review a PR”, “how to write my reports”) that the agent activates when the task calls for it. In the Skills tab:

  • Installed: your Skills, both the ones that ship with the app and your own.
  • Market (skills.sh): search the public Skills catalog and install with one click.

The Create skill button opens a form that saves it already in the right format:

  1. Name (e.g. “PR reviewer”).
  2. Scope: Global (available in all your projects) or Project (only in the current workspace).
  3. Description — important: it defines when the Skill activates (“use it when the user asks to review a pull request…”).
  4. Instructions in Markdown: the step-by-step the AI should follow.
Creating a Skill: name, scope, when it activates and its instructions.

MCP (Model Context Protocol) is the standard for giving the agent external tools: databases, APIs, services. In the MCP tab:

  • Installed: your configured MCP servers, with hot reload (“Recargar servidores MCP” — Reload MCP servers).
  • Marketplace: a built-in catalog for discovering MCP servers with no API key needed — it combines a curated selection with the official public registry.
  • Add MCP by hand: paste the server command (e.g. -y @scope/server) and choose the scope to save it in.
The MCP marketplace: search, install and reload servers without touching config files.

Some remote MCP servers ask you to sign in with your account on the service. Their card shows a Sign in button that opens the browser so you can authorize access. From then on:

  • The session keeps itself alive. When the token expires, the app renews it automatically and repeats the request that had failed, without you noticing. The card says so with “Session active; it renews itself.”. Only if the renewal fails (for example, because you revoked access from your account on the service) will you need to click Sign in again.
  • Open chats pick up the new session. If you sign in again, chats that were already open start using the new token on their next request: there’s no need to open another chat. The server is also relaunched on its own, so the card stops showing in red.
  • A single authorized application. Context Code reuses the registration it already made on the service instead of creating a new application every time you sign in, so duplicate entries don’t pile up in your account settings on the service.

Servers launched with the python command now start on Windows even if python isn’t on the PATH: the app uses the Windows py launcher (the one the official Python installer sets up) or, if it isn’t there, the Python managed by uv. The Microsoft Store python.exe shortcut, which actually opens the Store, is ignored. Thanks to this, these servers no longer show up in red just because of how Python is invoked.

If there really is no Python on the computer, or the server is missing a package, the card explains what to install and offers Retry. See Troubleshooting.

The catalog has been cleaned up: servers whose job Context Code already does natively were removed from the curated catalog and also from the public registry search:

Removed serverReplaced by
filesystemThe agent’s native file tools (read, write, edit and search)
gitThe Git panel and the built-in git tools
fetchThe built-in web tools (WebFetch and WebSearch)
memoryThe assistant’s memory
sequential-thinkingThe built-in step-by-step reasoning tool (Think)
postgresThe SQL console and the built-in database MCP
puppeteerPlaywright and Chrome DevTools, which come built in
codebase-memoryThe project’s code graph

If you already had one installed, it isn’t uninstalled automatically: under Installed it shows the “Superseded” status, in amber —not red, because it isn’t an error—, with an explanation of which built-in feature replaces it and a note that you can uninstall it because it’s no longer needed. Removing it or keeping it is up to you.

The Plugins tab lists larger packages that combine several capabilities (for example, web scraping plugins with Puppeteer/Playwright, database administration, CI/CD automation…). You enable and disable them right from the list.

The LSP tab manages the language servers that power autocomplete and diagnostics in the editor: see which ones you have, install them and update them.

The Skills active in the chat are listed with their toggle: they step in on their own when needed.
The installed MCP servers and the tools each one contributes.
A plugin is a GitHub repo: it brings Skills and MCP servers together, and the engine loads them on its own.