Gemini CLI MCP Server

A lightweight FastMCP server that wraps Google's Gemini CLI as MCP tools. Works with any MCP client — Hermes Agent, Claude Code, Claude Desktop, Cursor, etc.

Prerequisites

• Node.js — for the Gemini CLI
• Python 3.11+ — for the MCP server
• Gemini CLI — installed and authenticated

npm install -g @google/gemini-cli
gemini --version          # Verify install
gemini                    # Run once to complete OAuth login

Installation

# Clone the repo
git clone https://github.com/jxsprt/gemini-mcp-server.git
cd gemini-mcp-server

# Create venv and install deps
python3 -m venv .venv
.venv/bin/pip install -r requirements.txt

Configuration

Add to your MCP client's config:

Hermes Agent (~/.hermes/config.yaml)

mcp_servers:
  gemini-cli:
    command: "/path/to/gemini-mcp-server/.venv/bin/python3"
    args: ["/path/to/gemini-mcp-server/server.py"]
    timeout: 240

Restart your Hermes gateway. Tools will be available as mcp_gemini_cli_*.

Claude Desktop (claude_desktop_config.json)

{
  "mcpServers": {
    "gemini-cli": {
      "command": "/path/to/gemini-mcp-server/.venv/bin/python3",
      "args": ["/path/to/gemini-mcp-server/server.py"]
    }
  }
}

Claude Code

fastmcp install claude-code /path/to/gemini-mcp-server/server.py

Cursor

fastmcp install cursor /path/to/gemini-mcp-server/server.py -e .

Tools

gemini_prompt

Send any prompt to Gemini CLI non-interactively. Supports model selection and JSON output.

Safe by default — uses --approval-mode auto-edit (auto-approves file edits, prompts for shell). Pass dangerous=true for full yolo mode.

gemini_prompt(prompt="Explain TCP handshake", model="gemini-2.5-flash")
gemini_prompt(prompt="Refactor this module", dangerous=True)     # full auto-approval

gemini_plan

Read-only audit and review mode. Uses --approval-mode plan — guarantees no file mutations or shell execution. Safe for whole-repo analysis and code reviews.

gemini_plan(prompt="Review this auth module for security issues")

gemini_version

Returns the installed Gemini CLI version.

gemini_version()  # → "0.41.2"

Verification

cd /path/to/gemini-mcp-server
.venv/bin/python -c "
import server
print('Version:', server.gemini_version())
print('Prompt:', server.gemini_prompt('Say hello in one word'))
print('Plan:', server.gemini_plan('What tools does this server have?'))
"

How It Works

The server shells out to gemini -p "prompt" --approval-mode yolo for each tool call. It does NOT use the Google Gen AI Python SDK — it goes through the Gemini CLI binary (OAuth-authenticated).

Key design decisions:

• Stdio transport — runs as a subprocess of your MCP client
• No hardcoded paths — discovers gemini on PATH
• No API keys in config — uses the CLI's existing OAuth session
• Retry-friendly — generous timeouts (120s default, 240s for plan mode) to handle Gemini's capacity retries

Notes

• Capacity errors: Gemini's reasoning models can hit rate limits. The CLI retries up to 7 times with backoff. If it fails, try again later or use a different model.
• Plan mode is safe: --approval-mode plan explicitly prevents the agent from writing files or executing shell commands. Read-only, guaranteed.
• 1M token context: Gemini CLI supports the full 1M token context window of Gemini models.

License

MIT