Skip to content

tutorial.md

Latest commit

 

History

History
111 lines (71 loc) · 3.5 KB

tutorial.md

File metadata and controls

111 lines (71 loc) · 3.5 KB

Tutorial: Getting started with gemini-plugin

This tutorial walks you through installing the plugin, running your first validation, and seeing the auto-trigger hooks in action. By the end, you will have Gemini actively reviewing your work inside Claude Code.

Prerequisites

  • Claude Code installed and running
  • A Google AI Studio API key (free at aistudio.google.com/app/apikey)
  • uv installed (provides uvx for running the MCP server)
  • jq installed (used by hook scripts)

Step 1: Install the plugin

Open Claude Code and add the SynthForge marketplace, then install the plugin:

/plugin marketplace add azmym/SynthForge
/plugin install gemini-plugin@synthforge

Step 2: Enter your API key

During installation, Claude Code prompts you for your Google AI Studio API key:

Gemini API Key: ________

The input is masked and the key is stored securely in your system keychain (not in settings files or plain text). You only enter this once; it persists across sessions.

If you don't have a key yet, get one free at aistudio.google.com/app/apikey.

Step 3: Verify the installation

Start a new Claude Code session. You should see the gemini MCP server connected:

claude mcp list
# Expected: gemini: ... - Connected

Check that skills are available:

/gemini-plugin:gemini-validate

You should see the slash command auto-complete.

Step 4: Run your first validation

Ask Claude to create a plan for any task:

Plan how to add a login page to this project

When Claude finishes the plan and exits plan mode, the ExitPlanMode hook fires automatically. Gemini's validator reviews the plan and either approves it (you proceed normally) or blocks with specific gaps to address.

Step 5: Try manual validation

You can validate anything on demand:

/gemini-plugin:gemini-validate Is my authentication flow handling token refresh correctly?

The validator spawns, calls Gemini, and returns a structured verdict with gaps, hallucinations (if any), and recommended next actions.

Step 6: Try grounded research

Ask a question that benefits from live web data:

/gemini-plugin:gemini-research What is the latest stable version of Next.js?

The researcher returns an answer with citations and a freshness timestamp.

Step 7: Enable brainstorming mode (optional)

If you're in a design session and want every prompt grounded automatically:

/gemini-plugin:gemini-brainstorm-on

Now every prompt you type gets search-grounded by Gemini before Claude answers. Turn it off when done:

/gemini-plugin:gemini-brainstorm-off

What happens next

With the plugin active, six hooks fire automatically at key moments:

  1. Session start builds a risk map of your repo
  2. User prompt grounds post-cutoff questions (keyword-gated, or always-on in brainstorm mode)
  3. Plan complete validates plans before you see them
  4. Destructive Bash challenges dangerous commands before execution
  5. Pre-compact preserves session state before context compaction
  6. Stop audits "done" claims before Claude stops working

All of these block Claude when Gemini finds issues, surfacing the critique inline so you never miss it.

Next steps