---

name: openmed description: "Route OpenMed local-first clinical NLP, PHI de-identification, healthcare interoperability, model runtime, multimodal intake, and privacy-risk workflows." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

OpenMed repo skill

Use this skill when a task asks to build, debug, or plan with OpenMed, the local-first Python/Swift/Android/browser SDK for clinical extraction, de-identification, document intake, model runtimes, and healthcare data handoffs. Keep examples synthetic unless the user explicitly confirms a local, authorized PHI boundary.

Quick orientation

• Package/distribution: openmed; import root: openmed.
• Python requirement: Python 3.10+.
• Core install: pip install openmed.
• Common editable install for a checkout: uv pip install -e ".[dev]" or a targeted editable install such as pip install -e ".[cli,service]".
• The CLI entry point is openmed; run openmed --help to list commands.
• Read references/repo-provenance.md before deciding whether this skill is current for a checkout.

Minimal import check

import openmed
print(openmed.__version__)
print(openmed.list_models(include_remote=False)[:3])

For no-download tests, use a fixture loader instead of a real model. The bundled scripts/openmed_quickstart_smoke.py shows this pattern for analyze_text and deidentify.

Route map

Task family	Read this
Extract, mask, replace, hash, date-shift, audit, stream, or batch PHI/PII	sub-skills/deidentification-privacy/SKILL.md
Run clinical/biomedical NER, section/context assertions, labs, meds, timelines, grounding, or codeable concepts	sub-skills/clinical-extraction-grounding/SKILL.md
De-identify or evaluate tabular/structured releases, quasi-identifiers, k-anonymity, l-diversity, DP, leakage gates, or compliance evidence	sub-skills/structured-risk-evaluation/SKILL.md
Use the openmed CLI, REST/gRPC/MCP surfaces, clients, adapters, FHIR/OMOP/HL7/OpenMRS/DHIS2/OpenHIM, or framework connectors	sub-skills/interoperability-serving/SKILL.md
Select/cache/prefetch models, inspect ModelLoader, choose CPU/CUDA/MLX/CoreML/ONNX/Torch, export for mobile/browser, or debug runtime parity	sub-skills/model-runtimes-mobile/SKILL.md
Extract text/layout/metadata from PDFs, Office files, Markdown, images/OCR, DICOM, SMS/chatlogs, calendars, contacts, or verify redaction fidelity	sub-skills/multimodal-document-intake/SKILL.md

Shared references and scripts

• references/package-map.md — read for the package/module map, CLI families, and optional extras by workflow.
• references/install-and-configuration.md — read when choosing extras, offline/cache settings, privacy-safe configuration, or environment checks.
• references/troubleshooting.md — read for cross-cutting install/import, optional dependency, privacy, backend, and stale-skill failures.
• scripts/check_openmed_environment.py — run to inspect installation, CLI, and optional backend availability without invoking models.
• scripts/openmed_quickstart_smoke.py — run for a deterministic synthetic analyze_text + deidentify smoke that avoids network/model downloads.

Core safety rules

• Do not paste real PHI into remote agent prompts, logs, code comments, golden files, or issue text. Use synthetic examples in all generated code.
• OpenMed is local-first after required artifacts are available; model downloads, remote adapters, telemetry-enabled paths, and user integrations are separate boundaries that must be explicitly accepted.
• Do not bundle restricted terminology or DUA-gated assets. UMLS, SNOMED CT, CPT, MIMIC, i2b2, n2c2, and similar assets must be user-supplied or bridged out-of-process.
• Aggregate F1 is not enough for privacy work: direct identifiers, critical leakage, span integrity, multilingual IDs, date shifts, surrogate consistency, and quantized recall deltas need separate checks.
• If the current checkout commit, package version, public APIs, CLI commands, or major docs differ from references/repo-provenance.md, refresh this skill.