GC-101 · Module 3

Extensions & Agent Skills

3 min read

Extensions are community-built packages that bundle tools, commands, and configuration into installable units. Think of them as plugins for Gemini CLI. Each extension has a gemini-extension.json manifest that declares what it provides. The Extensions Gallery at geminicli.com is the central directory for discovering and installing extensions.

# Install an extension
gemini extensions install firebase

# List installed extensions
gemini extensions list

# Remove an extension
gemini extensions remove firebase

Notable community extensions include Firebase (deploy, manage, query Firestore), Cloud Run (container deployment), BigQuery (query execution and schema exploration), Data Commons (access Google's public datasets), and Google Workspace (interact with Docs, Sheets, and Drive). These extensions are particularly powerful because they give Gemini CLI native access to Google's ecosystem — a natural fit given the underlying Gemini models.

Agent skills are a preview feature that enables multi-step autonomous workflows. Unlike commands (which template a single prompt), skills define a sequence of actions the AI can take autonomously. The Jules extension is a prominent example — it creates a remote AI worker that can take on tasks asynchronously, working on a separate branch and reporting back when done.

The Jules extension deserves special attention. It connects Gemini CLI to Google's Jules service — an asynchronous AI agent that works on your codebase in the cloud. You assign Jules a task, it creates a branch, works through the problem, and presents a pull request when done. This is the "fire and forget" pattern: delegate routine work to Jules while you focus on high-judgment tasks.

Do This

  • Check the Extensions Gallery before building custom integrations
  • Use Jules for well-defined, self-contained tasks you'd delegate to a junior dev
  • Review Jules' pull requests as carefully as you would any human contributor

Avoid This

  • Install every extension you find — each one adds context weight
  • Use Jules for ambiguous, architecture-level decisions that need human judgment
  • Merge Jules' PRs without review — autonomous doesn't mean infallible