NanoBanana MCP

import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js"; import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js"; import { z } from "zod"; import axios from "axios"; import * as fs from "fs"; import * as path from "path"; import * as os from "os"; import * as dotenv from "dotenv"; // Load environment variables dotenv.config(); // User credentials from environment variables const API_KEY = process.env.GOOGLE_API_KEY || ''; const PROJECT_NUMBER = process.env.GOOGLE_PROJECT_NUMBER || ''; // Default output directory - saves to current working directory (where Claude Desktop is running) // This means images are saved in the project you have open in your IDE const DEFAULT_OUTPUT_DIR = process.env.OUTPUT_DIR || path.join(process.cwd(), 'nanobanana-images'); // Create the MCP server const server = new McpServer({ name: "Nano Banana", version: "1.0.0", }); // Helper to find the best image model async function getBestImageModel(): Promise<string> { try { const listResponse = await axios.get( `https://generativelanguage.googleapis.com/v1beta/models?key=${API_KEY}` ); const models = listResponse.data.models.map((m: any) => m.name); // Prefer imagen-4.0-fast, then imagen-3.0, then any imagen const preferred = [ 'models/imagen-4.0-fast-generate-001', 'models/imagen-3.0-generate-001', ]; for (const pref of preferred) { if (models.includes(pref)) return pref; } // Fallback search const found = models.find((m: string) => m.includes('imagen') && m.includes('generate')); if (found) return found; // Default fallback if listing fails or returns nothing useful but we hope it works return 'models/imagen-3.0-generate-001'; } catch (error) { console.error("Error listing models, using default:", error); return 'models/imagen-3.0-generate-001'; } } server.tool( "generate_image", "Generate an image using Google's Imagen model and automatically save it", { prompt: z.string().describe("The text prompt to generate an image from"), outputDir: z.string().optional().describe("Directory to save the images. If not specified, images are saved to ./nanobanana-images in your current project directory"), aspectRatio: z.string().optional().describe("Aspect ratio of the image (e.g., '1:1', '16:9'). Default is '1:1'"), numberOfImages: z.number().optional().default(1).describe("Number of images to generate (1-4)"), }, async ({ prompt, outputDir, aspectRatio, numberOfImages }) => { const modelName = await getBestImageModel(); console.error(`Using model: ${modelName}`); // Construct request // We try the 'predict' endpoint first as it's standard for Imagen on Vertex/GenAI // But for API Key access, sometimes it's 'generateContent' on Generative Language API. // Based on previous search, Imagen 3 is on Vertex AI but also accessible via Gemini API Key? // Let's try the 'predict' schema specifically for Imagen. const url = `https://generativelanguage.googleapis.com/v1beta/${modelName}:predict?key=${API_KEY}`; const payload = { instances: [ { prompt: prompt } ], parameters: { sampleCount: numberOfImages || 1, // aspectRatio: aspectRatio || "1:1" // Some models take aspectRatio, some don't in 'parameters'. // For Imagen 3, it's usually in 'parameters' } }; if (aspectRatio) { (payload.parameters as any).aspectRatio = aspectRatio; } try { const response = await axios.post(url, payload); const predictions = response.data.predictions; if (!predictions || predictions.length === 0) { return { content: [ { type: "text", text: "No images generated." } ] }; } // Ensure output directory exists const saveDir = outputDir || DEFAULT_OUTPUT_DIR; if (!fs.existsSync(saveDir)) { fs.mkdirSync(saveDir, { recursive: true }); } const contentItems: any[] = []; const savedFiles: string[] = []; const timestamp = new Date().toISOString().replace(/[:.]/g, '-'); for (let i = 0; i < predictions.length; i++) { const prediction = predictions[i]; let base64Image = ""; let mimeType = "image/png"; if (prediction.bytesBase64Encoded) { base64Image = prediction.bytesBase64Encoded; } else if (prediction.mimeType && prediction.bytesBase64Encoded) { mimeType = prediction.mimeType; base64Image = prediction.bytesBase64Encoded; } else { contentItems.push({ type: "text", text: `Unknown prediction format: ${JSON.stringify(prediction).substring(0, 100)}...` }); continue; } // Save image to file const extension = mimeType.includes('png') ? 'png' : 'jpg'; const filename = `nanobanana_${timestamp}_${i + 1}.${extension}`; const filepath = path.join(saveDir, filename); const buffer = Buffer.from(base64Image, 'base64'); fs.writeFileSync(filepath, buffer); savedFiles.push(filepath); contentItems.push({ type: "image", data: base64Image, mimeType: mimeType }); } // Return summary with file paths const summaryText = `Generated ${predictions.length} image(s) using ${modelName}\n\nSaved to:\n${savedFiles.map(f => `- ${f}`).join('\n')}`; contentItems.unshift({ type: "text", text: summaryText }); return { content: contentItems }; } catch (error: any) { // Fallback to generateContent if predict fails (maybe it's a Gemini model that can do images?) // But we are explicitly using an 'imagen' model name. const errorMessage = error.response ? JSON.stringify(error.response.data) : error.message; console.error("Image generation failed:", errorMessage); return { content: [ { type: "text", text: `Error generating image: ${errorMessage}` } ], isError: true, }; } } ); async function main() { const transport = new StdioServerTransport(); await server.connect(transport); console.error("Nano Banana MCP Server running on stdio"); } main();