openmed

OpenMed is an open-source, local-first healthcare AI toolkit that turns clinical text into structured, de-identified insights on hardware you control. It provides clinical named entity recognition (NER) and PII detection/de-identification that runs 100% on-device after required model artifacts are available, with no cloud dependency and no patient data leaving your network. Key capabilities: - **Clinical NER**: Extract diseases, drugs, anatomy, genes, and more using curated models (e.g., `disease_detection_superclinical`, `pharma_detection_superclinical`). - **PII de-identification**: Detect and redact names, dates, SSNs, phone numbers, emails, addresses, and national IDs with methods like mask, replace, hash, and date shifting. Supports smart entity merging and policy-aware pipelines (HIPAA/GDPR/research profiles). - **Multilingual support**: 38 PII language routes (35 model-backed), including English, Spanish, French, German, Italian, Portuguese, Dutch, Arabic, Hindi, Japanese, Turkish, and more. - **On-device execution**: Runs on CPU, CUDA, Apple Silicon (MLX), Android (ONNX Runtime Mobile), and browsers (Transformers.js/WebGPU). - **Multiple interfaces**: Python API, REST service (FastAPI), batch processing, command-line interface, MCP server, and Swift/OpenMedKit for iOS. - **Agent skills**: Portable Agent Skills for Claude Code, OpenAI Codex, OpenCode, and compatible agents. - **Air-gapped support**: Load models from local directories without contacting Hugging Face Hub. OpenMed includes a curated model catalog with over 2,200 manifest entries, and the SDK source is Apache-2.0 licensed. It is designed for healthcare workflows where privacy and data control are critical, with deployment owners responsible for validating model terms and clinical fitness.