The tools follow the study-series-instance hierarchy, which means the assistant can start from a patient id and end at pixel data without you assembling UIDs by hand. The field selector on the instance query is the one to remember: instance metadata is large, and asking for specific tags keeps the response readable.
A Python MCP server that exposes DICOM operations as tools: searching a PACS, drilling into a study, and retrieving instances locally.
- `query_studies` — search by `patient_id`, `study_date`, `accession_number`, `patient_name` or additional filters
- `query_series` — the series inside a study, by `study_instance_uid`
- `query_instances` — instance metadata within a series, with `fields_to_retrieve` so you pull only the tags you need
- `move_dicom_entity_to_local_server` — move a study, series or single instance to the local server
- `get_local_instance_pixel_data` — read pixel data for an instance you already moved locally
`pip install -e .`, plus a reachable DICOM server. The endpoint is named twice, differently — `DICOM_SERVER_BASE_URL` in the main file and `DICOM_SERVER_URL` in the client's env block — so set both to the same address rather than picking one. Documentation is in Spanish.
One command — pip install -e .