Replicate Flux MCP

English | 中文

Replicate Flux MCP is an advanced Model Context Protocol (MCP) server that empowers AI assistants to generate high-quality images and vector graphics. By default it uses black-forest-labs/flux-schnell for raster images and recraft-ai/recraft-v3-svg for SVG output. You can override the curated image/SVG models through environment variables, and image tools also accept a per-call model_id override from the built-in allowlist.

📑 Table of Contents

• Getting Started & Integration
  • Setup Process
  • Cursor Integration
  • Claude Desktop Integration
  • Smithery Integration
• Glama.ai Integration
• Codex Integration
• Features
• Documentation
  • Available Tools
  • Available Resources
  • Available Prompts
  • Structured Output
  • Environment Variables
• Development
  • Testing
• Technical Details
• Troubleshooting
• Contributing
• License
• Resources
• Examples

🚀 Getting Started & Integration

Setup Process

1. Obtain a Replicate API Token

   • Sign up at Replicate
   • Create an API token in your account settings

2. Choose Your Integration Method

   • Follow one of the integration options below based on your preferred MCP client

3. Ask Your AI Assistant to Generate an Image

   • Simply ask naturally: "Can you generate an image of a serene mountain landscape at sunset?"
   • Or be more specific: "Please create an image showing a peaceful mountain scene with a lake reflecting the sunset colors in the foreground"

4. Explore Advanced Features

   • Try different parameter settings for customized results
   • Experiment with SVG generation using generate_svg
   • Use batch image generation or variant generation features

Cursor Integration

Method 1: Using mcp.json

1. Create or edit the .cursor/mcp.json file in your project directory:

{
  "mcpServers": {
    "replicate-flux-mcp": {
      "command": "env REPLICATE_API_TOKEN=YOUR_TOKEN npx",
      "args": ["-y", "replicate-flux-mcp"]
    }
  }
}

2. Replace YOUR_TOKEN with your actual Replicate API token
3. Restart Cursor to apply the changes

Method 2: Manual Mode

1. Open Cursor and go to Settings
2. Navigate to the "MCP" or "Model Context Protocol" section
3. Click "Add Server" or equivalent
4. Enter the following command in the appropriate field:

env REPLICATE_API_TOKEN=YOUR_TOKEN npx -y replicate-flux-mcp

5. Replace YOUR_TOKEN with your actual Replicate API token
6. Save the settings and restart Cursor if necessary

Claude Desktop Integration

1. Create or edit the mcp.json file in your configuration directory:

{
  "mcpServers": {
    "replicate-flux-mcp": {
      "command": "npx",
      "args": ["-y", "replicate-flux-mcp"],
      "env": {
        "REPLICATE_API_TOKEN": "YOUR TOKEN"
      }
    }
  }
}

2. Replace YOUR_TOKEN with your actual Replicate API token
3. Restart Claude Desktop to apply the changes

Smithery Integration

This MCP server is available as a hosted service on Smithery, allowing you to use it without setting up your own server.

1. Visit Smithery and create an account if you don't have one
2. Navigate to the Replicate Flux MCP server page
3. Click "Add to Workspace" to add the server to your Smithery workspace
4. Configure your MCP client (Cursor, Claude Desktop, etc.) to use your Smithery workspace URL

For more information on using Smithery with your MCP clients, visit the Smithery documentation.

Glama.ai Integration

This MCP server is also available as a hosted service on Glama.ai, providing another option to use it without local setup.

1. Visit Glama.ai and create an account if you don't have one
2. Go to the Replicate Flux MCP server page
3. Click "Install Server" to add the server to your workspace
4. Configure your MCP client to use your Glama.ai workspace

For more information, visit the Glama.ai MCP servers documentation.

Codex Integration

Add the server to ~/.codex/config.toml:

[mcp_servers.replicate]
command = "npx"
args = ["-y", "replicate-flux-mcp"]
env = { REPLICATE_API_TOKEN = "your-replicate-api-token", REPLICATE_IMAGE_MODEL_ID = "your-image-model-id", REPLICATE_SVG_MODEL_ID = "your-svg-model-id" }
startup_timeout_sec = 30_000

Replace the env values as needed. If you omit REPLICATE_IMAGE_MODEL_ID / REPLICATE_SVG_MODEL_ID, the server uses black-forest-labs/flux-schnell for images and recraft-ai/recraft-v3-svg for SVGs.

Curated tools validate model overrides against built-in allowlists:

• Image generation: black-forest-labs/flux-schnell, google/imagen-4, black-forest-labs/flux-kontext-pro, ideogram-ai/ideogram-v3-turbo, black-forest-labs/flux-1.1-pro, black-forest-labs/flux-dev
• SVG generation: recraft-ai/recraft-v3-svg
• Legacy run_model extras: minimax/video-01, luma/reframe-video, topazlabs/video-upscale, topazlabs/image-upscale, szcho/codeformer, tencentarc/gfpgan

🌟 Features

• 🖼️ High-Quality Image Generation — Flux Schnell raster images by default, with environment-variable and per-tool image model overrides.
• 🎨 Vector Graphics — Recraft V3 SVG for logos, icons, and diagrams.
• 📊 Batch + Variants — Generate N images from N prompts or N variants of one prompt (seed-based or prompt-modifier-based).
• 🧩 Arbitrary Replicate Models — run_replicate_model escape hatch accepts any owner/name[:version] reference, with get_model_schema introspection for the OpenAPI input schema. Optional allowlist via REPLICATE_MODEL_ALLOWLIST.
• 📦 Structured Output — Every generate_* tool returns machine-readable structuredContent alongside human-readable content, matching a per-tool outputSchema (URL, prompt, format, aspect ratio, per-variant seed, etc).
• ⏳ Progress Notifications — Batch and variant generation emit notifications/progress for clients that opt in via progressToken, so long runs aren't black-boxed.
• 💬 Curated Prompts — 5 ready-made prompt templates (logo, portrait, svg-icon, product-shot, isometric-diagram) surfaced in Claude Desktop's slash palette and Cursor's @-menu.
• 🏷️ Proper Tool Annotations — readOnlyHint / destructiveHint / openWorldHint / idempotentHint set correctly so clients can reason about safety and cost.
• 🪵 Structured Logging — Server-side errors travel over notifications/message instead of stderr.
• 🔌 Universal MCP Compatibility — MCP protocol 2025-11-25; works with Claude Desktop, Cursor, Cline, Zed, and any spec-compliant client.
• 🔍 Generation History — Browse past runs through imagelist, svglist, and predictionlist resources.

📚 Documentation

Available Tools

generate_image

Generates an image based on a text prompt using the configured image model (allowlist only).

{
  prompt: string;                // Required: Text description of the image to generate
  model_id?: string;             // Optional: Override image model (allowlist only)
  seed?: number;                 // Optional: Random seed for reproducible generation
  go_fast?: boolean;             // Optional: Run faster predictions with optimized model (default: true)
  megapixels?: "1" | "0.25";     // Optional: Image resolution (default: "1")
  num_outputs?: number;          // Optional: Number of images to generate (1-4) (default: 1)
  aspect_ratio?: string;         // Optional: Aspect ratio (e.g., "16:9", "4:3") (default: "1:1")
  output_format?: string;        // Optional: Output format ("webp", "jpg", "png") (default: "webp")
  output_quality?: number;       // Optional: Image quality (0-100) (default: 80)
  num_inference_steps?: number;  // Optional: Number of denoising steps (1-4) (default: 4)
  disable_safety_checker?: boolean; // Optional: Disable safety filter (default: false)
  support_image_mcp_response_type?: boolean; // Optional: Return embedded image content when supported (default: true)
}

generate_multiple_images

Generates multiple images based on an array of prompts using the configured image model (allowlist only).

{
  prompts: string[];             // Required: Array of text descriptions for images to generate (1-10 prompts)
  model_id?: string;             // Optional: Override image model (allowlist only)
  seed?: number;                 // Optional: Random seed for reproducible generation
  go_fast?: boolean;             // Optional: Run faster predictions with optimized model (default: true)
  megapixels?: "1" | "0.25";     // Optional: Image resolution (default: "1")
  aspect_ratio?: string;         // Optional: Aspect ratio (e.g., "16:9", "4:3") (default: "1:1")
  output_format?: string;        // Optional: Output format ("webp", "jpg", "png") (default: "webp")
  output_quality?: number;       // Optional: Image quality (0-100) (default: 80)
  num_inference_steps?: number;  // Optional: Number of denoising steps (1-4) (default: 4)
  disable_safety_checker?: boolean; // Optional: Disable safety filter (default: false)
  support_image_mcp_response_type?: boolean; // Optional: Return embedded image content when supported (default: true)
}

generate_image_variants

Generates multiple variants of the same image from a single prompt using the configured image model (allowlist only).

{
  prompt: string;                // Required: Text description for the image to generate variants of
  model_id?: string;             // Optional: Override image model (allowlist only)
  num_variants: number;          // Required: Number of image variants to generate (2-10, default: 4)
  prompt_variations?: string[];  // Optional: List of prompt modifiers to apply to variants (e.g., ["in watercolor style", "in oil painting style"])
  variation_mode?: "append" | "replace"; // Optional: How to apply variations - 'append' adds to base prompt, 'replace' uses variations directly (default: "append")
  seed?: number;                 // Optional: Base random seed. Each variant will use seed+variant_index
  go_fast?: boolean;             // Optional: Run faster predictions with optimized model (default: true)
  megapixels?: "1" | "0.25";     // Optional: Image resolution (default: "1")
  aspect_ratio?: string;         // Optional: Aspect ratio (e.g., "16:9", "4:3") (default: "1:1")
  output_format?: string;        // Optional: Output format ("webp", "jpg", "png") (default: "webp")
  output_quality?: number;       // Optional: Image quality (0-100) (default: 80)
  num_inference_steps?: number;  // Optional: Number of denoising steps (1-4) (default: 4)
  disable_safety_checker?: boolean; // Optional: Disable safety filter (default: false)
  support_image_mcp_response_type?: boolean; // Optional: Return embedded image content when supported (default: true)
}

generate_svg

Generates SVG/vector output based on a text prompt using the configured SVG model (allowlist only).

{
  prompt: string;                // Required: Text description of the SVG to generate
  size?: string;                 // Optional: Size of the generated SVG (default: "1024x1024")
  style?: string;                // Optional: Style of the generated image (default: "any")
                                // Options: "any", "engraving", "line_art", "line_circuit", "linocut"
}

prediction_list

Retrieves a list of your recent predictions from Replicate.

{
  limit?: number;  // Optional: Maximum number of predictions to return (1-100) (default: 50)
}

get_prediction

Gets detailed information about a specific prediction.

{
  predictionId: string;  // Required: ID of the prediction to retrieve
}

run_model

Runs a whitelisted Replicate model with a raw input payload (useful for video or restoration models).

{
  model_id: string;                // Required: Replicate model id (allowlist only)
  input?: Record<string, unknown>; // Optional: Raw input payload for the model
}

run_replicate_model

Runs any model hosted on Replicate by its owner/name[:version] reference. Use this as an escape hatch when none of the curated tools fit. Call get_model_schema first if you don't know the input shape.

{
  model: string;                              // Required: 'owner/name' or 'owner/name:version'
  input: Record<string, unknown>;             // Required: Model input parameters
  prefer_wait?: number;                       // Optional: Seconds to block waiting for sync output (1-60, default 60)
  return_as?: "url" | "base64" | "both";      // Optional: How to return file outputs (default "url")
}

Set the REPLICATE_MODEL_ALLOWLIST env var (comma-separated owner/name entries) to restrict which models can be invoked. Unset = any model allowed. Set-but-empty = deny all (the server fails closed rather than silently allowing everything).

get_model_schema

Fetches the OpenAPI input schema and description for a Replicate model so you can pass the right parameters to run_replicate_model.

{
  model: string;  // Required: Replicate model reference in 'owner/name' form
}

Available Resources

imagelist

Browse your history of generated images created with the configured image model.

svglist

Browse your history of generated SVG outputs created with the configured SVG model.

predictionlist

Browse all your Replicate predictions history.

Available Prompts

Curated templates surfaced in Claude Desktop's slash menu and Cursor's @-palette. Each one fills in sensible defaults then delegates to the relevant generation tool.

Structured Output

Every generate_* tool returns both human-readable content (text + image blocks) and machine-readable structuredContent that matches the tool's outputSchema.

Clients that understand MCP structured output can consume URLs and metadata directly without parsing prose.

Environment Variables

💻 Development

1. Clone the repository:

git clone https://github.com/awkoy/replicate-flux-mcp.git
cd replicate-flux-mcp

2. Install dependencies:

npm install

3. Start the TypeScript watcher:

npm run watch

4. Build the project:

npm run build

5. Smoke-test the server with the MCP Inspector:

npm run inspector

6. Connect to Client:

{
  "mcpServers": {
    "image-generation-mcp": {
      "command": "npx",
      "args": [
        "/Users/{USERNAME}/{PATH_TO}/replicate-flux-mcp/build/index.js"
      ],
      "env": {
        "REPLICATE_API_TOKEN": "YOUR REPLICATE API TOKEN"
      }
    }
  }
}

Testing

This project currently has no automated test suite. Verification is done via:

• npm run build — TypeScript type-checking catches most regressions.
• npm run inspector — drives the built binary through the official MCP Inspector for end-to-end smoke testing of tools, resources, and prompts.

Contributions adding a proper test framework (e.g. Vitest + an MCP stdio client harness) are welcome.

⚙️ Technical Details

Stack

• Model Context Protocol SDK - Core MCP functionality for tool and resource management
• Replicate API - Provides access to state-of-the-art AI image generation models
• TypeScript - Ensures type safety and leverages modern JavaScript features
• Zod - Implements runtime type validation for robust API interactions

Configuration

The server can be configured by modifying the CONFIG object in src/config/index.ts or by setting the REPLICATE_IMAGE_MODEL_ID / REPLICATE_SVG_MODEL_ID environment variables to override the defaults:

export const CONFIG = {
  serverName: "replicate-flux-mcp",
  serverVersion: "0.4.0",
  imageModelId: process.env.REPLICATE_IMAGE_MODEL_ID ?? "black-forest-labs/flux-schnell",
  svgModelId: process.env.REPLICATE_SVG_MODEL_ID ?? "recraft-ai/recraft-v3-svg",
  pollingAttempts: 25,
  pollingInterval: 2000, // ms
  modelAllowlistConfigured: process.env.REPLICATE_MODEL_ALLOWLIST !== undefined,
  modelAllowlist: (process.env.REPLICATE_MODEL_ALLOWLIST ?? "")
    .split(",")
    .map((s) => s.trim())
    .filter(Boolean),
};

Switching models (no code changes)

Use env vars when launching the server (works with npx, Cursor, Claude Desktop, etc.). Curated image/SVG tool overrides must be in the built-in allowlists:

# Stay on defaults:
REPLICATE_API_TOKEN=YOUR_TOKEN npx -y replicate-flux-mcp

# Switch to other allowlisted models
REPLICATE_IMAGE_MODEL_ID="google/imagen-4" \
REPLICATE_SVG_MODEL_ID="recraft-ai/recraft-v3-svg" \
REPLICATE_API_TOKEN=YOUR_TOKEN \
npx -y replicate-flux-mcp

modelAllowlist is evaluated once at process start from REPLICATE_MODEL_ALLOWLIST. Restart the server after changing it.

🔍 Troubleshooting

Common Issues

Authentication Error

• Ensure your REPLICATE_API_TOKEN is correctly set in the environment
• Verify your token is valid by testing it with the Replicate API directly

Safety Filter Triggered

• The model has a built-in safety filter that may block certain prompts
• Try modifying your prompt to avoid potentially problematic content

Timeout Error

• For larger images or busy servers, you might need to increase pollingAttempts or pollingInterval in the configuration
• Default settings should work for most use cases

🤝 Contributing

Contributions are welcome! Please follow these steps to contribute:

1. Fork the repository
2. Create your feature branch (git checkout -b feature/amazing-feature)
3. Commit your changes (git commit -m 'Add some amazing feature')
4. Push to the branch (git push origin feature/amazing-feature)
5. Open a Pull Request

For feature requests or bug reports, please create a GitHub issue. If you like this project, consider starring the repository!

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

🔗 Resources

• Model Context Protocol Documentation
• Replicate API Documentation
• Try for Free Collection
• Flux Schnell Model
• Recraft V3 SVG Model
• MCP TypeScript SDK
• Smithery Documentation
• Glama.ai MCP Servers

🎨 Examples

Multiple Prompts	Prompt Variants

Here are some examples of how to use the tools:

Batch Image Generation with generate_multiple_images

Create multiple distinct images at once with different prompts:

{
  "prompts": [
    "A red sports car on a mountain road", 
    "A blue sports car on a beach", 
    "A vintage sports car in a city street"
  ]
}

Image Variants with generate_image_variants

Create different interpretations of the same concept using seeds:

{
  "prompt": "A futuristic city skyline at night",
  "num_variants": 4,
  "seed": 42
}

Or explore style variations with prompt modifiers:

{
  "prompt": "A character portrait",
  "prompt_variations": [
    "in anime style", 
    "in watercolor style", 
    "in oil painting style", 
    "as a 3D render"
  ]
}

Made with ❤️ by Yaroslav Boiko