FLUX MCP Server & CLI

A Model Context Protocol (MCP) server and command-line tool for generating images using FLUX.1-dev with automatic model unloading to save VRAM and power.

Features

• 🎨 High-Quality Image Generation - Uses FLUX.1-dev for state-of-the-art image synthesis

• ⚡ Lazy Loading - Model loads only when needed

• 🔄 Auto-Unload - Automatically unloads model after configurable inactivity period (MCP mode)

• 💾 Memory Efficient - Uses bfloat16 for optimal VRAM usage (~12GB)

• 🎲 Reproducible - Seed-based generation for consistent results

• 📊 Status Monitoring - Check model status and VRAM usage

• 🔧 Runtime Configuration - Adjust timeout without restarting

• 🖥️ Dual Interface - Use via Claude Desktop (MCP) or command-line (CLI)

Requirements

• Python 3.10+

• NVIDIA GPU with 16GB+ VRAM (tested on RTX 4070 Ti Super)

• CUDA toolkit installed

• PyTorch with CUDA support

Installation

1. Clone the repository (or navigate to the project directory):

2. Install with UV (recommended):

Or install with pip:

3. Configure environment variables:

Configuration Options

Edit .env to customize:

MCP Server Registration

Add the server to your Claude Desktop configuration file:

macOS: ~/Library/Application Support/Claude/claude_desktop_config.json Windows: %APPDATA%\Claude\claude_desktop_config.json Linux: ~/.config/Claude/claude_desktop_config.json

Or if installed globally with pip:

After adding the configuration, restart Claude Desktop.

CLI Usage

In addition to the MCP server for Claude Desktop, you can use FLUX directly from the command line for completely offline and private image generation.

Quick Start

Generate Command

The main command for image generation:

Options:

• --steps, -s INTEGER - Number of inference steps (default: 28)

• --guidance, -g FLOAT - Guidance scale (default: 3.5)

• --width, -w INTEGER - Image width in pixels, must be multiple of 8 (default: 1024)

• --height, -h INTEGER - Image height in pixels, must be multiple of 8 (default: 1024)

• --seed INTEGER - Random seed for reproducibility

• --output, -o PATH - Custom output path (default: auto-generated)

• --output-dir PATH - Override output directory

• --interactive, -i - Interactive mode

• --verbose, -v - Verbose output with debug info

Examples:

Interactive Mode

Interactive mode allows you to generate multiple images without reloading the model:

Interactive workflow:

1. Enter your prompt

2. Configure parameters (steps, guidance, dimensions, seed)

3. Image generates and saves

4. Choose to generate another or exit

5. Model stays loaded between generations for faster subsequent images

Other Commands

Status Command:

Shows:

• Model information

• Output directory

• CUDA availability

• GPU name and VRAM usage

• Model cache location

Config Command:

Displays current configuration from environment variables.

Open Output:

Opens the output directory in your file manager (Linux: xdg-open, macOS: open, Windows: explorer).

Output Files

Generated images are saved with metadata:

• Image: flux_YYYYMMDD_HHMMSS_SEED.png

• Metadata: flux_YYYYMMDD_HHMMSS_SEED.json

Metadata JSON contains:

CLI vs MCP Server

CLI Mode:

• ✓ Completely offline and private (no Claude Desktop needed)

• ✓ Direct control from terminal

• ✓ Batch generation with interactive mode

• ✓ No auto-unload (process terminates after generation)

• ✓ Saves metadata JSON files

• ✓ Rich terminal UI with progress bars

MCP Server Mode:

• ✓ Integrated with Claude Desktop

• ✓ Natural language interface

• ✓ Auto-unload after timeout (saves power)

• ✓ Persistent background process

• ✓ Access from Claude conversations

Both modes share the same configuration, model cache, and output directory.

MCP Server Tools (Claude Desktop)

1. generate_image

Generate an image from a text prompt.

Parameters:

• prompt (required): Text description of the image

• steps (optional): Number of inference steps (default: 28, range: 20-50)

• guidance_scale (optional): Guidance scale (default: 3.5, range: 1.0-10.0)

• width (optional): Image width in pixels (default: 1024, range: 256-2048)

• height (optional): Image height in pixels (default: 1024, range: 256-2048)

• seed (optional): Random seed for reproducibility (random if not provided)

Example Usage in Claude:

2. unload_model

Immediately unload the FLUX model from GPU memory.

Example Usage:

3. get_status

Check the current status of the FLUX generator.

Returns:

• Model load status

• Time remaining until auto-unload

• Current VRAM usage

• Last access time

Example Usage:

4. set_timeout

Change the auto-unload timeout at runtime.

Parameters:

• timeout_seconds (required): New timeout in seconds (0 to disable)

Example Usage:

Usage Examples

Basic Image Generation

The server will:

1. Load the FLUX model (if not already loaded)

2. Generate the image

3. Save it to the output directory as YYYYMMDD_HHMMSS_{seed}.png

4. Return the file path, seed, and generation settings

5. Schedule auto-unload after 5 minutes (default)

Reproducible Generation

To generate the same image again, use the seed from a previous generation:

Custom Parameters

Memory Management

Check current status:

Manually unload to free VRAM:

Adjust auto-unload timeout:

How It Works

Auto-Unload Mechanism

1. Lazy Loading: The model is NOT loaded when the server starts

2. On-Demand Loading: Model loads automatically on first generation request

3. Timer Reset: Each generation resets the auto-unload timer

4. Automatic Cleanup: After the configured timeout with no activity:

   • Model is removed from memory

   • GPU cache is cleared (torch.cuda.empty_cache())

   • Python garbage collection runs

5. Seamless Reload: Model automatically reloads on next request

Memory Management

The server uses several strategies to minimize VRAM usage:

• bfloat16 precision instead of float32 (saves ~50% VRAM)

• Explicit cache clearing when unloading

• Threading for non-blocking auto-unload

• Lock-based synchronization for thread-safe operation

Output Files

Generated images are saved as:

Example: 20250126_143052_42.png

Troubleshooting

CUDA Out of Memory

Problem: Error during generation: "CUDA out of memory"

Note: The generator automatically uses sequential CPU offloading to reduce VRAM usage from ~28GB to ~12GB. This should work on 16GB GPUs like RTX 4070 Ti Super.

If you still get OOM errors:

1. Close other GPU applications:

   # Check what's using VRAM nvidia-smi

2. Reduce image dimensions:

   flux generate "prompt" --width 768 --height 768 # Or even smaller flux generate "prompt" --width 512 --height 512

3. Reduce inference steps:

   flux generate "prompt" --steps 20 # Default is 28

4. Restart the process if VRAM isn't fully freed:

   # CLI: Just run again (process exits after generation) # MCP: Restart Claude Desktop

Model Download Issues

Problem: Model download fails or times out

Solutions:

1. Check internet connection

2. Set a custom cache directory with more space:

   FLUX_MODEL_CACHE=/path/to/large/disk/cache

3. Download manually with HuggingFace CLI:

   huggingface-cli download black-forest-labs/FLUX.1-dev

Server Not Responding

Problem: Claude Desktop doesn't see the tools

Solutions:

1. Check Claude Desktop logs for errors

2. Verify the configuration path is absolute

3. Ensure UV is in PATH or use full path to UV binary

4. Restart Claude Desktop after config changes

5. Test the server manually:

   cd /path/to/flux-mcp uv run flux-mcp

Slow Generation

Problem: Image generation takes too long

Solutions:

1. Reduce steps parameter (try 20-25 instead of 28)

2. Ensure GPU is being used (check with nvidia-smi)

3. Close background applications to free GPU resources

4. Check that CUDA is properly installed

Permission Errors

Problem: Cannot write to output directory

Solutions:

1. Check directory permissions

2. Set a different output directory in .env:

   FLUX_OUTPUT_DIR=/home/$USER/flux_output

3. Create the directory manually:

   mkdir -p ~/flux_output chmod 755 ~/flux_output

Advanced Configuration

Custom Model Cache

To share the model cache across multiple projects or save space:

Disable Auto-Unload

To keep the model loaded permanently (uses more power but faster):

Or at runtime:

Logging

The server logs to stderr. To capture logs:

Performance Tips

Optimal Settings for RTX 4070 Ti Super (16GB)

• Resolution: Up to 1024x1024 comfortably

• Steps: 25-30 for good quality

• Batch size: 1 (model doesn't support batching well)

• Timeout: 300s for occasional use, 600s for active sessions

Generation Time Expectations

• 1024x1024, 28 steps: ~20-40 seconds (depending on prompt complexity)

• 512x512, 20 steps: ~5-10 seconds

• First generation: +10-15 seconds for model loading

Technical Details

Architecture

Key Components

• FluxGenerator: Manages model lifecycle, threading, and GPU memory (shared between CLI and MCP)

• Config: Loads environment variables and provides defaults (shared)

• MCP Server: Exposes tools via Model Context Protocol for Claude Desktop

• CLI Tool: Direct command-line interface for offline usage

Thread Safety

The generator uses a threading lock (threading.Lock) to ensure:

• Only one generation at a time

• Safe model loading/unloading

• No race conditions with auto-unload timer

License

MIT License - see LICENSE file for details

Contributing

Contributions welcome! Please:

1. Fork the repository

2. Create a feature branch

3. Make your changes

4. Submit a pull request

Support

For issues and questions:

• Check the Troubleshooting section above

• Review server logs for errors

• Open an issue on GitHub

Changelog

v0.1.0 (2025-01-26)

• Initial release

• FLUX.1-dev integration

• Auto-unload functionality (MCP mode)

• Four MCP tools (generate, unload, status, set_timeout)

• CLI tool with interactive mode (flux command)

• Shared architecture between CLI and MCP server

• Comprehensive documentation with CLI and MCP usage examples