Ask an assistant (MCP)

bioconice-mcp gives an MCP-capable assistant read-only SQL access to the public release. It runs on your machine, with no account and no token.

bioconice-mcp is a Model Context Protocol server. It gives an AI assistant the same read-only access as a DuckDB ATTACH on the release. The assistant can list tables, read column docs, run guarded SQL and adapt worked recipes.

The server runs locally over stdio. Your assistant starts it as a subprocess, and queries run on your machine, in DuckDB, against the public https://data.bioconice.cancerdatasci.org/latest/ release. No hosted endpoint is offered: bioc-on-ice publishes data and does not run queries for you.

You need uv installed. It provides uvx.

Install

Claude Code

claude mcp add bioconice -- uvx --from git+https://github.com/seandavi/bioc-on-ice bioconice-mcp

This writes to Claude Code’s config only (~/.claude.json). Add --scope project to write the project’s .mcp.json instead, so everyone working in that repository gets the server.

Claude Desktop

Add the server to claude_desktop_config.json (Settings → Developer → Edit Config; on macOS, ~/Library/Application Support/Claude/claude_desktop_config.json):

{"mcpServers": {"bioconice": {"command": "/absolute/path/to/uvx", "args": ["--from", "git+https://github.com/seandavi/bioc-on-ice", "bioconice-mcp"]}}}

Use the path that which uvx prints for command. Desktop is a GUI app that does not inherit your shell’s PATH, so a bare uvx fails with spawn uvx ENOENT. Quit Desktop fully and reopen it to load the change. The server’s log is ~/Library/Logs/Claude/mcp-server-bioconice.log.

Claude Code and Claude Desktop keep separate configs. Adding the server to one does not add it to the other.

Any other MCP client

Any client that launches stdio servers can run the same command:

uvx --from git+https://github.com/seandavi/bioc-on-ice bioconice-mcp

To pin a dated release instead of latest/, set BIOCONICE_RELEASE_URL=https://data.bioconice.cancerdatasci.org/v2026-11-01 in the server’s environment (see Versions and citation).

Tools

The assistant reads these descriptions itself. This list is for you.

Tool	What it returns
list_tables(namespace="")	Every table in the release, or one namespace’s: namespace, table, comment, row count and business key. The assistant calls it first.
describe_table(table)	One table’s columns: type and doc from the release’s manifest.json, plus the Bioregistry identifier prefix, coordinate convention and business-key membership declared in biocOnIce’s schemas. table is namespace.table, e.g. annotation.ncbi__gene.
resolve_release(...)	A release id and its valid_from/valid_to predicate, from a release id, a source version (ensembl=, icite=, ncbi_date=, or source= with source_version=). An unknown version errors with the list of known ones. The release holds current rows only, so a past release’s predicate returns only rows still current today.
query(sql, limit=1000, current=False)	columns, rows, row_count and truncated. Only one SELECT or WITH statement is accepted. Write statements, ATTACH, COPY, INSTALL/LOAD, PRAGMA, SET and a second statement are all rejected before DuckDB sees them. limit defaults to 1000 and must be 1–10000 (anything else is an error); a result with more rows than limit comes back truncated, with truncated: true. current=True keeps only current rows, so the SELECT list must include valid_to. Tables are named b.<namespace>.<table>.
recipes()	Worked SQL for common joins, each run against live data: gene → citing literature, datasets by cell-type rollup (a recursive is_a closure), gene models at a release, one taxon’s coverage.

It also serves one resource, bioconice://docs/{table}: the describe_table output as Markdown, plus the ATTACH line.

Example prompts

• “Using bioconice, how many papers cite the PubMed papers NCBI links to human TP53?”
• “Which human blood datasets carry T cells or any T cell subtype? Walk the Cell Ontology, don’t list terms by hand.”
• “Which tables have Ensembl gene coordinates, and are they 0- or 1-based?”