IDC agent skill

The Imaging Data Commons Skill is an Agent Skill: a packaged set of instructions, reference guides, and helper scripts that an AI assistant loads on demand when you ask it about cancer imaging data.

It is not Claude-specific. It follows the open Agent Skills format, so it works with Claude (claude.ai, Claude Desktop, Claude Code, API) and with any other agent that supports the format — or you can load its instruction file into any assistant's context by hand.

Unlike the hosted server, the skill has your assistant write and run Python directly against the idc-index package in a code-execution sandbox. There is no IDC server in the loop, and no network round trip for metadata queries — but it only works where your assistant can execute code.

What it gives your assistant

• A grounded understanding of the IDC data model and the idc-index metadata tables — which table answers which question, and how they join.

• Validated query patterns, so it writes SQL that runs instead of SQL that looks plausible.

• Download via the idc-index Python API and the idc command-line tool, including manifests and directory templates.

• Viewer URLs, so you can eyeball a series in the browser before committing to a download.

• License checks (CC BY vs CC BY-NC) and ready-to-use citations in several formats.

Reference guides it bundles

The skill ships a set of topic guides that the assistant pulls in only when relevant, covering:

• the idc-index tables, their join keys, and when to use each

• end-to-end workflows, from a research question to downloaded files

• SQL patterns for cohort building and aggregation

• clinical (non-imaging) data — discovering tables, and joining them to imaging

• direct access to the public S3/GCS buckets, and reading DICOM straight from cloud storage

• DICOMweb access

• digital pathology and slide microscopy

• querying the IDC BigQuery tables

• the idc command-line interface

• reading the underlying Parquet index directly

This is the main reason to add the skill. Several of these surfaces — BigQuery, DICOMweb, direct bucket and Parquet access, the pathology and clinical-data workflows — are not exposed by the hosted server at all.

Requirements

• Your assistant must be able to execute Python (3.10 or newer).

• Network access to PyPI, to install idc-index, and to the public IDC buckets.

• No authentication for IDC data — it is all public.

• The BigQuery and Google Healthcare DICOMweb paths additionally need your own GCP credentials. Everything else works without a cloud account.

Do I need this if I already use the hosted server?

They overlap, and running both is the recommended setup.

• The hosted server covers discovery, cohort building, SQL, retrieval, licenses, and citations, with no setup beyond a URL.

• The skill adds the surfaces listed above, and does its metadata work locally rather than over the network.

• If your assistant can run Python, there is no downside to having both installed — the skill's own setup guide walks through adding the server alongside it.

If you would rather see the three access paths side by side, see Using IDC with an AI assistant.

Staying current

The skill is versioned, and each release is pinned to a specific idc-index release and a specific IDC data version. IDC publishes new data releases several times a year, so a stale skill will describe a stale IDC.

• Watch the releases page and update when a new version lands.

• After updating, start a new conversation — an assistant binds its skills at the start of a session.

• To check what your assistant is actually working from, ask it which IDC data version it sees.

Where it lives

The skill is developed in the open at ImagingDataCommons/imaging-data-commons-skill and released under the MIT license (the data it reaches carries its own per-collection licenses). If the assistant gives you an incomplete or incorrect answer, please report it via the repository's issue tracker — the issue template is designed for exactly that.

Ready to install it? See Setting up the skill.