Claude Gemini Plugin (claude-gemini-plugin)
Supercharge your Claude Code experience with the power of Google Gemini.
claude-gemini-plugin is a high-performance integration for Claude Code that delegates complex tasks, deep code reviews, and adversarial testing to Google's Gemini models. By combining Claude's exceptional coding ability with Gemini's massive context window and advanced reasoning, you get the best of both worlds in your terminal.
🚀 Key Features
- Advanced Code Review: Deeply analyze your working tree or branch diffs with
/gemini:review.
- Adversarial Testing: Challenge your own design choices with
/gemini:adversarial-review to find edge cases and potential flaws.
- Autonomous Task Rescue: Delegate implementation tasks or bug investigations to Gemini with
/gemini:rescue (or the gemini:gemini-rescue subagent when Claude routes work to Gemini proactively).
- Background Jobs: Run long-running investigations with
--background, then monitor via /gemini:status, fetch with /gemini:result, or stop them with /gemini:cancel.
- Model Selection: Choose your preferred Gemini model from the CLI or set a persistent default via
/gemini:setup --model. If you do not pass --model, the plugin defaults to auto-gemini-3 (an alias that routes to the best Gemini 3 series model; auto-gemini-2.5, pro, flash, and flash-lite are also supported).
- Reasoning Control: Request a t-shirt-sized thinking level with
--thinking <off|low|medium|high> (default: medium). Caveat: the companion currently parses --thinking but does not yet ship a per-invocation thinking override to the Gemini runtime — the flag is a one-shot/one-time hint only. For persistent reasoning levels, edit settings.json and rerun /gemini:setup.
- Streamed or Compact Output: Pass
--stream-output for raw model and thought chunks, or accept the default which surfaces compact markers like [tool], [thinking], and live progress to stderr.
- Stop-time Review Gate (optional): Automatically request a Gemini review at the end of each Claude turn.
- Job Health & Event Tail:
/gemini:status shows live health classifications and a recent-events / last-event tail so you can tell active runs from possibly_stalled ones at a glance.
🛠 Installation
-
Install Gemini CLI:
Make sure the Google Gemini CLI is on your PATH. Run !gemini from inside Claude Code to launch the OAuth flow, or set GEMINI_API_KEY.
npm install -g @google/gemini-cli
If Gemini CLI is missing when you run /gemini:setup, the plugin will offer to install it for you.
-
Add the marketplace and install:
/plugin marketplace add seungmanchoi/claude-gemini-plugin
/plugin install gemini@claude-gemini-plugin
-
Initialize Setup:
/gemini:setup
/gemini:setup --enable-review-gate
/gemini:setup --disable-review-gate
This verifies the CLI version, checks authentication, and shows the active model and review-gate state.
Note: Gemini 3.x models require @google/gemini-cli >= 0.33.0. /gemini:setup will flag a mismatch and suggest either an upgrade or falling back to --model auto-gemini-2.5.
📖 Commands
/gemini:setup
Check whether Gemini CLI is installed, configure the default model, and toggle the stop-time review gate. If Gemini is missing, this command will offer to install it for you.
/gemini:setup
/gemini:setup --enable-review-gate
/gemini:setup --disable-review-gate
If Gemini is installed but not authenticated, run !gemini from inside Claude Code to authenticate interactively.
/gemini:review
Performs a comprehensive code review of working-tree or branch changes.
/gemini:review [--base <ref>] [--scope <auto|working-tree|branch>] [--model <modelId>] [--thinking <off|low|medium|high>] [--stream-output]
/gemini:adversarial-review
A steerable review mode that actively looks for flaws in your implementation. It uses the same review target selection as /gemini:review, so you can run e.g. /gemini:adversarial-review --base main challenge whether this was the right caching and retry design to focus the critique.
/gemini:adversarial-review [focus area] [--base <ref>] [--scope <auto|working-tree|branch>] [--model <modelId>] [--thinking <off|low|medium|high>] [--stream-output]
