Connecting Gemini to Claude Code via MCP — Using Two AIs Without Token Anxiety

At some point while using Claude Code, you'll run into this message.

"Approaching context limit"

Large file analysis, long log parsing, repetitive processing — Claude is powerful, but it burns through tokens fast. And switching to Gemini in a separate browser tab for copy-pasting defeats the purpose of having an AI assistant in the first place.

This post covers how to connect Gemini as an MCP tool inside Claude Code, so you can say "handle this one with Gemini" and have it just work.

What is MCP?

MCP (Model Context Protocol) is the protocol that lets Claude Code call external tools and services. Think of it as a plugin system for Claude Code.

Once you register an MCP server, Claude can call its functions directly mid-conversation. Wrap the Gemini API in an MCP server, and Gemini becomes a tool Claude can use on demand.

User → Claude Code → (when needed) Gemini MCP Server → Gemini API

Prerequisites

Node.js 20 or higher
Claude Code CLI installed
A Gemini API key from Google AI Studio

Step 1 — Create the MCP Server Directory

Placing the server inside the Claude config folder makes it available across all projects.

mkdir -p ~/.claude/mcp-servers/gemini
cd ~/.claude/mcp-servers/gemini

Step 2 — Write package.json

{
  "name": "gemini-mcp-server",
  "version": "1.0.0",
  "type": "module",
  "main": "index.js",
  "dependencies": {
    "@google/generative-ai": "^0.24.0",
    "@modelcontextprotocol/sdk": "^1.10.2"
  }
}

"type": "module" is required for ES Module syntax.

Step 3 — Write the MCP Server (index.js)

This server exposes two tools:

ask_gemini — sends a text prompt
ask_gemini_with_file — sends a file's contents along with the prompt (for large file analysis)

import { Server } from "@modelcontextprotocol/sdk/server/index.js";
import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js";
import {
  CallToolRequestSchema,
  ListToolsRequestSchema,
} from "@modelcontextprotocol/sdk/types.js";
import { GoogleGenerativeAI } from "@google/generative-ai";
import fs from "fs";

const apiKey = process.env.GEMINI_API_KEY;
if (!apiKey) {
  console.error("GEMINI_API_KEY environment variable is not set.");
  process.exit(1);
}

const genAI = new GoogleGenerativeAI(apiKey);

const server = new Server(
  { name: "gemini-mcp-server", version: "1.0.0" },
  { capabilities: { tools: {} } }
);

// Tool definitions
server.setRequestHandler(ListToolsRequestSchema, async () => ({
  tools: [
    {
      name: "ask_gemini",
      description:
        "Send a question or task to Gemini AI. Best used for long document analysis, large file summarization, and repetitive text processing that would otherwise burn through Claude's context.",
      inputSchema: {
        type: "object",
        properties: {
          prompt: { type: "string", description: "The prompt to send to Gemini" },
          model: {
            type: "string",
            description: "Model to use (default: gemini-2.0-flash)",
            default: "gemini-2.0-flash",
          },
        },
        required: ["prompt"],
      },
    },
    {
      name: "ask_gemini_with_file",
      description:
        "Send a file's contents along with a prompt to Gemini. Ideal for analyzing large files.",
      inputSchema: {
        type: "object",
        properties: {
          prompt: { type: "string", description: "The prompt to send alongside the file" },
          file_path: { type: "string", description: "Absolute path to the file to analyze" },
          model: {
            type: "string",
            description: "Model to use (default: gemini-2.0-flash)",
            default: "gemini-2.0-flash",
          },
        },
        required: ["prompt", "file_path"],
      },
    },
  ],
}));

// Tool execution handler
server.setRequestHandler(CallToolRequestSchema, async (request) => {
  const { name, arguments: args } = request.params;

  try {
    if (name === "ask_gemini") {
      const model = genAI.getGenerativeModel({
        model: args.model || "gemini-2.0-flash",
      });
      const result = await model.generateContent(args.prompt);
      return { content: [{ type: "text", text: result.response.text() }] };
    }

    if (name === "ask_gemini_with_file") {
      const fileContent = fs.readFileSync(args.file_path, "utf-8");
      const model = genAI.getGenerativeModel({
        model: args.model || "gemini-2.0-flash",
      });
      const prompt = `${args.prompt}\n\n--- File contents (${args.file_path}) ---\n${fileContent}`;
      const result = await model.generateContent(prompt);
      return { content: [{ type: "text", text: result.response.text() }] };
    }
  } catch (error) {
    return {
      content: [{ type: "text", text: `Error: ${error.message}` }],
      isError: true,
    };
  }
});

const transport = new StdioServerTransport();
await server.connect(transport);

Step 4 — Install Dependencies

npm install

Step 5 — Register the MCP Server with Claude Code

Important: Do NOT manually add mcpServers to ~/.claude/settings.json — it will fail schema validation. Always use the claude mcp add command.

claude mcp add gemini -s user \
  -e GEMINI_API_KEY=your_api_key_here \
  -- node ~/.claude/mcp-servers/gemini/index.js

Option	Description
`-s user`	Register globally (available in all projects)
`-e KEY=VALUE`	Environment variable injected at server startup
`--`	Separator between CLI flags and the run command

Step 6 — Verify the Connection

claude mcp list

Expected output:

gemini: node ~/.claude/mcp-servers/gemini/index.js - ✓ Connected

Once you see ✓ Connected, restart Claude Code and Gemini is ready to use mid-conversation.

Usage Examples

Just tell Claude what you want in natural language:

Analyze this file with Gemini: /path/to/large-log.txt

Use Gemini to summarize this long document

This is repetitive text transformation — handle it with Gemini

Claude will decide whether to call ask_gemini or ask_gemini_with_file based on the request.

When to Use Gemini vs Claude

Task	Recommended
Large log / file analysis	Gemini (1M token context)
Long document summarization	Gemini
Repetitive text processing	Gemini
Code implementation / editing	Claude
Architecture design	Claude
Complex debugging	Claude

Troubleshooting

Added mcpServers to settings.json manually and got an error

Settings validation failed: Unrecognized field: mcpServers

Claude Code's settings.json schema does not allow mcpServers directly. Always use claude mcp add. The actual config is stored in ~/.claude.json.

429 Too Many Requests

RESOURCE_EXHAUSTED: Free tier quota exceeded

The Gemini API free tier has daily and per-minute limits. Options:

Wait until tomorrow — free tier resets daily
Issue a new API key — create a new project in Google AI Studio
Upgrade to paid — link billing in Google AI Studio

To swap API keys:

claude mcp remove gemini
claude mcp add gemini -s user -e GEMINI_API_KEY=new_key -- node ~/.claude/mcp-servers/gemini/index.js

Connected but tool calls aren't working

Fully restart the Claude Code session. Newly registered MCP servers aren't picked up until Claude Code restarts.

Summary

[1] Create directory
    mkdir -p ~/.claude/mcp-servers/gemini
            ↓
[2] Write package.json + index.js
            ↓
[3] npm install
            ↓
[4] Register with claude mcp add
    (never edit settings.json directly)
            ↓
[5] Verify with claude mcp list → ✓ Connected
            ↓
[6] Restart Claude Code and use it
    "Handle this one with Gemini"

Set it up once and Gemini is available in any project. Claude handles the precision work; Gemini handles the volume. Together, the token anxiety mostly goes away.