Dicomweb MCP Server

What PantelisGeorgiadis/dicomweb-mcp-server MCP server does

PantelisGeorgiadis/dicomweb-mcp-server MCP server exposes a DICOMweb archive to MCP-compatible clients. It provides a search-and-inspection workflow that starts with studies, narrows to series and instances, and then retrieves metadata, report text, or rendered frames from a selected instance.

The server expects a DICOMweb implementation that supports both QIDO-RS and WADO-RS. QIDO-RS endpoints are used for study, series, and instance searches. WADO-RS endpoints provide instance metadata, report content, and rendered frames. Orthanc is given as an example of a compatible archive, but the requirement is support for the documented DICOMweb endpoints.

How it works

Search tools accept a space-separated query made from key=value pairs. Keys can use DICOM keyword names, such as PatientName, or hexadecimal tags. Queries can also include result limits, offsets, fuzzy name matching, and requests for all available attributes. Wildcards are supported where the DICOMweb archive permits them.

Results are ordered according to the resource level: studies by study date with newer results first, series by series date with newer results first, and instances by ascending Instance Number. The returned identifiers can then be passed to tools that operate on a specific series or instance.

Structured Report and Encapsulated PDF discovery uses modality and SOP Class filtering within a study. Structured Reports are converted into readable text. Encapsulated PDF retrieval extracts text from searchable PDFs; scanned PDFs without an embedded text layer are not OCR-processed.

Setup and configuration

The server requires Node.js 20 or newer and a reachable DICOMweb archive. Install it with npx -y dicomweb-mcp-server, or install the package globally with npm and run the dicomweb-mcp-server command.

Set DICOMWEB_HOST to the archive’s base DICOMweb URL. Optional authentication settings support either basic or bearer: basic authentication uses DICOMWEB_USER and DICOMWEB_PASS, while bearer authentication uses DICOMWEB_TOKEN. DICOMWEB_TIMEOUT can set a request timeout in milliseconds; omitting it disables the timeout. Configuration can be supplied through a .env file in the server’s working directory or through the MCP client’s environment block.

The README includes stdio configurations for Claude Desktop and Cursor. The same server can also be registered in VS Code’s MCP configuration.

Tools and capabilities

The available tools are:

• Search studies, series, and instances using DICOMweb query filters.
• Locate Encapsulated PDF instances within a study.
• Extract text from an Encapsulated PDF DICOM instance.
• Locate Structured Report instances and convert them to readable text.
• Format all DICOM attributes from one instance without retrieving pixel data.
• Render a selected one-based frame as JPEG or PNG.

Frame rendering accepts frame number 1 for single-frame instances. The rendering tool returns the selected image as an inline image.

Limitations and notes

The archive must implement the required QIDO-RS and WADO-RS routes. Authentication is optional, but the configured credentials must match the selected authentication mode. PDF text extraction is limited to PDFs containing searchable text and does not perform OCR. Metadata inspection deliberately excludes pixel data, while image access is provided through the separate frame-rendering tool.

The workflow depends on identifiers returned by earlier searches: series searches require a study Instance UID, instance searches require both study and series UIDs, and instance-level retrieval tools require the relevant SOP Instance UID.