Unified computer vision and image processing server with automated GPU memory management and multi-provider AI routing.
omni-image-tools-mcp is a Python-based Model Context Protocol server created by Alex Santos. It enables AI coding agents to visually inspect layouts, extract OCR text, resize and compress graphics, and execute vision queries locally or via cloud models.
Seamless fallback between local Ollama vision models and high-throughput cloud endpoints.
Design choices, architectural patterns, and engineering decisions implemented for extreme reliability and developer experience.
Allows agents to route lightweight inspection tasks to local models (e.g. minicpm-v, llava) while delegating high-resolution queries to GPT-4o or Claude 3.5 Sonnet.
Monitors NVIDIA VRAM usage and cleans allocated PyTorch/CUDA memory buffers after compute-heavy operations to prevent out-of-memory crashes.
Powered by Pillow and OpenCV, offering batch resizing, thumbnail synthesis, EXIF sanitization, and color space normalizations.
Operate fully offline with Ollama vision models, guaranteeing complete data sovereignty for corporate environments.
Real-world usage, terminal configuration, and technical integration commands.
{
"mcpServers": {
"omni-image-tools": {
"command": "uvx",
"args": ["omni-image-tools-mcp"],
"env": {
"OLLAMA_HOST": "http://localhost:11434"
}
}
}
}Key clarifications regarding licensing, compatibility, deployment, and security.
Yes. All image processing operations fallback smoothly to multi-threaded CPU execution.
Ollama models such as minicpm-v:8b and llava-phi3 provide outstanding results for UI inspection.
Yes, both structured vision OCR and bounding-box text localization are supported.
Images are dynamically downscaled and tiled to match optimal neural input dimensions.