albumentations

This MCP server exposes Albumentations-based image augmentation through natural language prompts. You can load an image (or URL/base64 payload), describe the desired transformation in plain English, and the server will translate that prompt into an Albumentations pipeline to produce an augmented image. Core tools include loading images for processing, applying prompts to generate transformations, validating prompts to see the structured transforms without processing, and listing available transforms and presets. For example, prompts like "add blur and rotate 15 degrees" will map to a Gaussian blur followed by a rotation. You can also query quick references for transforms, set a default seed for reproducible results, and check the current pipeline status and output locations. If you pair this with VLM features, you can also explore image-to-image edits and Gemini/VLM-assisted refinements, though those require enabling VLM in the environment.

To use with a MCP client, start the server via the configured command and then issue API/tool calls such as load_image_for_processing, augment_image, validate_prompt, list_available_transforms, and get_pipeline_status. The flow typically involves loading an image (or providing a URL), then sending a prompt to augment_image and retrieving the resulting image and metadata. You can enumerate presets like segmentation, portrait, and lowlight to quickly apply common pipelines, or use get_quick_transform_reference for a compact keyword guide when prompting.

Prerequisites:

• Python 3.8+ (recommended)
• pip (Python package manager)

Install the MCP server package from PyPI:

pip install albumentations-mcp

Run the MCP server using UVX (as documented in the repo):

uvx albumentations-mcp

Optional: If you plan to use VLM features (Gemini/Nano Banana), install with VLM extras:

pip install 'albumentations-mcp[vlm]'

Configure environment variables for your setup if needed (examples):

export MCP_LOG_LEVEL=INFO
export OUTPUT_DIR=./outputs
export ENABLE_VLM=false

For editor integrations (Claude/KiRO), place or reference the MCP config in your editor’s settings as shown in the README examples.

Tips and notes:

• The server supports base64 payloads, remote URLs, and local file paths as image sources via load_image_for_processing.
• Use validate_prompt to preview the transformed transforms without generating images, which helps refine prompts.
• The OUTPUT_DIR path is where results and any intermediate data are stored; ensure the path exists or that the server has write permissions.
• If you enable ENABLE_VLM, you’ll need to configure VLM_PROVIDER and VLM_MODEL (or provide a config/vlm.json file as described in the README).
• When prompting, you can rely on presets such as segmentation, portrait, and lowlight for quick-start pipelines.
• For reproducible results, set DEFAULT_SEED; this seed is used to seed stochastic components of the pipeline.
• If you encounter issues, consult the troubleshooting resources mentioned in the documentation (e.g., get_getting_started_guide, troubleshooting_common_issues).