An experiment in agent-driven scientific data exploration, built around FinnGen and partner human-genetics datasets exposed through the FinnGenie results API (FinnGen R13 + UK Biobank + MVP meta-analyses, eQTL Catalogue R7, Open Targets 25.12, GTEx, Genebass, GenCC, Monarch -- 29 datasets at last count).
Same shape as jx -- a catalog of marimo notebooks plus thin operational skills -- with one difference: jx queries local DuckDB files, fgx hits a REST API. FinnGenie ships https://finngenie.fi/api/v1/* with bearer auth, predictable paths, and TSV by default. That collapses the data-access layer to httpx.get, so there is no Python SDK, no MCP server, no schema cache -- nb01's first cell shows the bare-curl equivalent of every API call below it. (The official genetics-mcp-server registers 29 tools that are all thin httpx wrappers around the same REST endpoints; none use MCP-specific primitives like sampling or elicitation, and a direct curl reproduces every result.)
A Python SDK would be lighter in syntax (finngenie.credible_sets_by_gene("PCSK9") vs three lines of httpx), but it adds a packaging layer that has to keep up with the API. The TSV-by-default response means duckdb's read_csv_auto over HTTPS turns "API access" into "SQL access" without an SDK -- we accept the extra httpx.get lines per notebook in exchange for not maintaining a dependency.
Each notebook ships with a committed session snapshot under notebooks/__marimo__/session/ so the molab preview renders cell outputs (plots, tables) without re-executing. Click a badge to view the rendered notebook in molab, or fork it from there.
| Notebook | Role | Preview |
|---|---|---|
nb01_pcsk9_walkthrough.py |
Gene -> credible sets -> lead variant -> colocalization |  |
nb02_variant_phewas.py |
Variant PheWAS across endpoints | |
nb03_phenotype_locus_zoom.py |
Phenotype-driven locus zoom | |
nb04_gene_exome_burden.py |
Gene-based exome burden results | |
- Create an API key at finngenie.broadinstitute.org (
MCP/API KEYS->Create key). The same key works for both the MCP and the REST API. cp .env.example .envand paste your key.- From a fresh Claude Code session in this clone, ask: help me get started. The
getting-startedskill installsuvandmarimo-pair, launches nb01 in a live marimo kernel, and hands off to interactive composition.
If you want to skip the agent-assist and just open the notebook: just notebook (uses uvx marimo edit --sandbox, provisions deps from the PEP 723 header on first launch).
Prerequisites: uv is the only thing you need to install; the notebook's PEP 723 header pulls in marimo, polars, httpx, altair, and python-dotenv. Optional: duckdb on $PATH if you want to run ad-hoc SQL against the same TSV endpoints from your shell.
fgx/
README.md # this file
Justfile # `just notebook`
.env.example # FINNGENIE_TOKEN= placeholder
.claude/skills/
getting-started/SKILL.md # cold-clone bootstrap
compose-notebook/SKILL.md # edit nb01 in place vs copy-fork to nb02_*
notebooks/
nb01_pcsk9_walkthrough.py # gene -> credible sets -> lead variant -> colocalization
nb02_variant_phewas.py # variant PheWAS across endpoints
nb03_phenotype_locus_zoom.py # phenotype-driven locus zoom
nb04_gene_exome_burden.py # gene-based exome burden results
__marimo__/session/ # committed session snapshots for molab previews
That's the whole repo. getting-started walks a fresh clone to a running marimo kernel; compose-notebook covers picking the right path (edit-in-place vs copy-fork) and the right endpoint when the user asks for a new genetics analysis. Both skills are intentionally thin -- as the catalog grows and cross-file function reuse appears, compose-notebook will need to grow into a per-module catalog table the way jx's compose-notebook does.
BSD 3-Clause. See LICENSE.
If fgx is useful in your work, please cite both fgx and the underlying FinnGenie service. CITATION.cff TBD.