GWAS Harmoniser MCP

An MCP server that lets AI agents harmonize local GWAS summary-statistics files with the strict workflow at gwasharmonizer.org.

Use it when a researcher asks to harmonize or harmonise GWAS sumstats, standardize columns, align effect and non-effect alleles, handle GRCh37/GRCh38 liftover, validate rsIDs, run harmonization QC, or prepare summary statistics for meta-analysis, downstream analysis, submission, or a paper. It does not run association tests.

Install

From the repository root:

scripts/install-gwasharmonizer-mcp

The installer registers the gwasharmonizer MCP server and installs the bundled gwas-harmonizer Agent Skill for Codex and Claude Code. It requires Python 3.12+, uv, and a reachable GWAS Harmoniser backend.

By default, local inputs must be under ~/Desktop. Files are uploaded to the configured live service for processing; raw variant rows are not returned to the agent through the MCP.

Agent tools

• start_harmonization: start harmonization for a local GWAS file. Use output_profile="strict_meta" for a full-file META_READY METAL export and download_kind="data" when only the .tsv.gz output should be downloaded.
• get_harmonization: check progress, required confirmations, blockers, and artifact paths.
• confirm_harmonization: resume only with scientific metadata the user verified from authoritative study documentation.

Example request: Harmonize ~/Desktop/study.tsv.gz for meta-analysis, verify META_READY, and download only the output file.

See the public agent and MCP guide for routing terms and scientific boundaries.

Resumable uploads (0.1.2)

Uploads use checksum-verified 8 MiB chunks when the backend supports them. Each request has at most three transient attempts. Calling start_harmonization again with the same file/options after a transfer failure resumes the retained transfer; repeated new whole-file transfers are unnecessary. Preview completion survives a backend restart and is polled independently of the upload connection. Changed options can reuse retained source bytes while keeping a separate preview/job. Older backends use the bounded multipart compatibility path. Update the installed client with uv tool install --force ./tools/gwasharmonizer-mcp from the repository root; already-running client processes need to finish/restart to load the update.