Hosted server (MCP)

The IDC MCP server lets LLM agents query IDC directly. It exposes IDC's capabilities as Model Context Protocol tools, so an agent such as Claude can find, describe, and download IDC data on your behalf, in plain conversation.

The Model Context Protocol is an open standard for connecting LLM applications to external tools and data sources. Any spec-conformant MCP client can use the IDC server.

The server is live in production at:

https://api.imaging.datacommons.cancer.gov/mcp

It is public and unauthenticated — no API key, account, or config file needed. Point a client at the URL and ask it to find an imaging cohort.

The MCP server and the REST API are released together as version 3, currently in beta. The contract may still change in response to feedback before the final 3.0.0 release. To see exactly which beta build is deployed, ask your agent to call the get_idc_version tool (or call the REST API's GET /v3/version) — the api_version field reports the running release. Feedback is welcome on the IDC support forum.

Shared concepts

The MCP server and the REST API are two surfaces over a single backend: a capability is implemented once and exposed in both. The difference is who is calling — you write the code against REST, whereas MCP is called by an LLM agent, with prescriptive tool descriptions and an idc://guide resource so the agent follows the recommended "ground-first" workflow on its own.

Because the two share a backend, the data model and query surfaces documented for the REST API apply unchanged to MCP.

For how this server compares with the IDC agent skill — the other way to give an assistant access to IDC — see Using IDC with an AI assistant.

In this section

• Connecting a client — set up claude.ai, Claude Desktop, or any other MCP client.

• Tools and resources — the full list of capabilities the server exposes.

The server is developed in the open at ImagingDataCommons/IDC-REST-MCP.