Gemini CLI Plugin for Claude Code
The Codex plugin showed that giving Claude Code a second brain to delegate to is genuinely useful — code reviews, rescue debugging, adversarial design challenges. Gemini is just as capable in that role, and sometimes better depending on the task. So why not have both?
This plugin brings Google's Gemini CLI into Claude Code as a peer. Same architecture as the Codex plugin, same patterns — different model behind the curtain. Sharing is caring.
And I could not help myself and sent Gemini to do Codex for Gemini so we can close the circle ... https://github.com/madciapka/codex-gemini-skill :)
Commands
| Command | Description |
|---|
/gemini:setup | Check if Gemini CLI is installed and authenticated |
/gemini:rescue | Delegate investigation, debugging, or implementation to Gemini |
/gemini:review | Run a code review against local git changes |
/gemini:adversarial-review | Challenge your design choices and assumptions |
/gemini:status | Show active and recent Gemini jobs |
/gemini:result | Show output from a completed job |
/gemini:cancel | Cancel a running background job |
Installation
Prerequisites
- Claude Code CLI installed
- Node.js 18+ (for the companion runtime script)
- Gemini CLI installed and authenticated
# Install Gemini CLI (pick one)
npm install -g @google/gemini-cli
brew install gemini-cli
# Authenticate (run interactively once)
gemini
Install the plugin
Option A: Install via marketplace (recommended)
From inside a Claude Code session, run:
/plugin marketplace add madciapka/gemini-claude-plugin
/plugin install gemini@google-gemini
/reload-plugins
/gemini:setup
Option B: Load directly (for development/testing)
git clone https://github.com/madciapka/gemini-claude-plugin.git
claude --plugin-dir ./gemini-claude-plugin/plugins/gemini
Use /reload-plugins inside Claude Code to pick up changes without restarting.
Usage
Delegate a task
/gemini:rescue fix the flaky test in src/auth.test.ts
Gemini gets full access to your working directory and can read files, run commands, and make edits.
Run a code review
/gemini:review
Reviews your current working tree changes (staged + unstaged). For branch reviews:
/gemini:review --base main
Adversarial review
/gemini:adversarial-review
/gemini:adversarial-review --base main focus on auth and data handling
Challenges your implementation approach — looks for violated invariants, missing failure paths, race conditions, and assumptions that break under stress.
Background execution
For long-running tasks, use --background:
/gemini:rescue --background refactor the payment module to use the new API
/gemini:status
/gemini:result
Flags
| Flag | Description |
|---|
--background | Run in the background |
--wait | Run in the foreground (default) |
--model <name> | Override the Gemini model |
--yolo | Auto-accept all Gemini actions |
--sandbox | Run Gemini in sandbox mode |
--base <ref> | Set the base branch for reviews |
--scope <mode> | Review scope: auto, working-tree, or branch |
Architecture
.claude-plugin/marketplace.json Marketplace catalog
plugins/gemini/
.claude-plugin/plugin.json Plugin metadata
agents/gemini-rescue.md Thin forwarder subagent
commands/*.md User-invocable slash commands
skills/*/SKILL.md Internal skills (prompting, runtime, result handling)
scripts/gemini-companion.mjs Main Node.js runtime
scripts/lib/*.mjs Library modules (CLI wrapper, git, state, rendering)
prompts/*.md Review prompt templates
The plugin follows a companion script pattern:
- Claude Code routes
/gemini:* commands to markdown-defined skills
- The
gemini-rescue subagent forwards tasks to the companion script
gemini-companion.mjs spawns the Gemini CLI in headless mode (gemini -p "..." -o text)- Output is returned verbatim to Claude Code — no rewriting or summarization
Background jobs are tracked via state files in .gemini-companion/ within the workspace.
Review Prompts
The plugin ships with two review prompt templates:
- Standard review (
prompts/review.md) — finds bugs, security issues, and quality problems
- Adversarial review (
prompts/adversarial-review.md) — actively tries to break confidence in the change, focusing on failure modes, trust boundaries, and rollback safety
Both enforce grounding rules to prevent hallucinated findings.
Why this exists
