Replicate Video AI MCP Server

A Model Context Protocol (MCP) server for generating videos using Replicate's video AI models.

Features

• Text-to-Video Generation: Create videos from text prompts
• Image-to-Video Generation: Animate still images with motion prompts
• Multiple Models: Support for Wan 2.2, Google Veo 3, and Kling 2.1
• Async Operations: Handle long-running video generation with status checking
• Terminal Mode: Built-in CLI for testing without MCP overhead

Supported Models

Alias	Model	Type	Features
wan-t2v-fast	Wan 2.2 Fast T2V	Text-to-Video	Fast (~30s), affordable
wan-i2v-fast	Wan 2.2 Fast I2V	Image-to-Video	Fast animation
veo3	Google Veo 3	Both	Premium quality with audio
kling-master	Kling 2.1 Master	Both	High quality, 5/10s duration

Setup

1. Set your Replicate API token:

export REPLICATE_API_TOKEN=your_token_here

Or create a .env file:

REPLICATE_API_TOKEN=your_token_here

2. Make the run script executable:

chmod +x run.sh

Usage

Terminal Mode

List available models:

./run.sh list-models

Generate text-to-video:

./run.sh t2v wan-t2v-fast "A sunset over the ocean"
./run.sh t2v veo3 "Dancing robot" --resolution 1080p
./run.sh t2v kling-master "City timelapse" --duration 10

Generate image-to-video:

./run.sh i2v wan-i2v-fast images/car.webp "Car driving forward"
./run.sh i2v veo3 images/test.jpg "Add rain effect"

Check generation status:

./run.sh continue <prediction_id>

Test async flow:

./run.sh test-async

Debug mode:

./run.sh debug t2v wan-t2v-fast "Test prompt"

MCP Server Mode

Start the MCP server:

./run.sh run

MCP Tools

generate_video_from_text

Generate a video from a text prompt.

Parameters:

• prompt (required): Text description of the video
• model: Model to use (default: wan-t2v-fast)
• resolution: Video resolution (480p, 720p, 1080p)
• aspect_ratio: Aspect ratio (16:9, 9:16, 1:1)
• duration: Duration in seconds (for Kling only)
• negative_prompt: What to avoid (for Veo3, Kling)

generate_video_from_image

Generate a video from an image with motion prompt.

Parameters:

• image_path (required): Path to input image
• prompt (required): How to animate the image
• model: Model to use (default: wan-i2v-fast)
• resolution: Video resolution
• duration: Duration (for Kling only)
• negative_prompt: What to avoid

continue_operation

Check status of async video generation.

Parameters:

• prediction_id (required): The prediction ID
• wait_time: How long to wait (5-60 seconds)

Output

Videos are saved to:

~/Library/Application Support/Savant/replicate_video_ai/<storage_id>/
├── video.mp4        # Generated video
├── metadata.yaml    # Generation parameters
└── input.jpg        # Input image (if I2V)

Environment Variables

• REPLICATE_API_TOKEN (required): Your Replicate API token
• REPLICATE_VIDEOS_ROOT_FOLDER: Custom output directory
• REPLICATE_VIDEO_DEBUG: Enable debug mode (true/false)
• REPLICATE_VIDEO_DEFAULT_TIMEOUT: Default timeout in seconds
• REPLICATE_VIDEO_POLL_INTERVAL: Status check interval

Development

Build the server:

./run.sh build

Run tests:

./run.sh test

License

MIT