ask-gemini
model selection [-m], sandbox [-s], and changeMode:boolean for providing edits
LABSCO SUMMARY
A second model's answer without leaving the first one's session.
What it removes is the copy-paste between assistants — and the context cost, since the files are read on Gemini's side rather than pulled into the conversation. When the reply is long enough to be a set of edits, it comes back in chunks the agent pulls in order instead of as one truncated block.
What it is
A bridge to Google's CLI rather than to an API: your prompt is handed to the CLI already signed in on the machine, with @ syntax for pulling whole files and directories into Gemini's large context. When the answer is a set of edits, change mode splits it into chunks fetched by cache key.
What you get
- A prompt answered by Gemini with the model named per call and an optional sandboxed run
- Files and directories addressed with @ inside the prompt, which is the point — a large codebase gets read by Gemini instead of loaded into the assistant you are talking to
- An edit-shaped response mode whose chunks are retrieved one at a time by cache key, so a long answer arrives whole rather than truncated
- A brainstorming call taking a domain, constraints, an idea count and a named methodology — SCAMPER, Design Thinking — with a feasibility analysis alongside the ideas if you ask for it
- An echo call and a help call for confirming the server answers before blaming the model
Requirements
Node.js v16.0.0 or higher, run over stdio from npm as gemini-mcp-tool. The real dependency is a Google CLI installed and signed in on the same machine — this server holds no key of its own. Which CLI depends on your account: Google retired the Gemini CLI on 2026-06-18 for free, AI Pro and AI Ultra tiers, and from that date the tool selects the Antigravity CLI, agy, automatically — install it and run it once to sign in. Enterprise, standard-licence and paid-API-key users keep the old CLI by setting GEMINI_MCP_BACKEND=gemini. AGY_CLI_PATH points at the binary when it is not on the server's PATH, and GEMINI_MCP_TIMEOUT caps a run in minutes, 45 by default. The agy backend is experimental: print mode is limited to one Flash model, and a requested model or sandbox that cannot be honoured produces a notice rather than silence.
Setup effort
One command — claude mcp add gemini-cli -- npx -y gemini-mcp-tool
USE CASES
01
Getting a large codebase read without filling your context
Reference the directory with @ in the prompt and let Gemini do the reading; what comes back is the analysis, not the files.
02
Taking a second opinion on a decision
Put the same question to a different model from inside the session you are already in, naming the model on the call.
03
Getting edits back in full
Ask in change mode and pull the chunks by cache key in order, rather than losing the tail of a long answer.
04
Opening up a problem before choosing an approach
Run the brainstorm with your domain and constraints under a named methodology, and ask for the feasibility read alongside the ideas.
05
Working out whether it is the server or the CLI that is broken
Echo a message to prove the transport, then read the help output — a failure at that point is the backend, not the wiring.
5 tools
Counted once with only startup setup. Other versions, modes or optional credentials may expose a different list.
brainstorm
Generate novel ideas with dynamic context gathering. --> Creative frameworks (SCAMPER, Design Thinking, etc.), domain context integration, idea clustering, feasibility analysis, and iterative refinement.
fetch-chunk
Retrieves cached chunks from a changeMode response. Use this to get subsequent chunks after receiving a partial changeMode response.
Help
receive help information
ping
Echo
BEFORE YOU INSTALL
What to know first
Tools you get
5 tools when we connected. The documentation's tool list is out of date: the tools are ask-gemini, brainstorm, fetch-chunk, ping and Help. sandbox-test is not a tool; sandbox is an option of ask-gemini.
Requires
npx on PATH
Where it runs
All 3 documented clients run the server on your machine.
CONTEXT COST
Light Standing cost
Rated by the size of the 5 tool definitions that ride along on every turn, not by how many there are
INSTALLATION
Connect your client
Server setup
Claude Code
TERMINAL
claude mcp add gemini-cli -- npx -y gemini-mcp-toolMCP Clients
Claude Desktop
CLAUDE DESKTOP · JSON
{
"mcpServers": {
"gemini-cli": {
"command": "npx",
"args": [
"-y",
"gemini-mcp-tool"
]
}
}
}Warp
WARP · JSON
{
"mcpServers": {
"gemini-cli": {
"command": "npx",
"args": [
"-y",
"gemini-mcp-tool"
],
"env": {},
"working_directory": null,
"start_on_launch": true
}
}
}
