Gigwa MCP Server

An MCP server that drives a local or remote Gigwa installation over its REST API. It lets an MCP client (Claude Desktop / Claude Code) run the whole genotyping workflow in plain language: connect → import genotype data & metadata → run QC and diversity analyses → audit databases for import artifacts. Built for genomic-resources teams and genebanks, but works with any Gigwa instance.

• Import DArTseq SNP/Silico xlsx reports (with correct 2-row genotype calling) or plain VCF, plus per-individual metadata.

• Analyse read-only: genotypes are pulled out of Gigwa and all statistics are computed in Python (scikit-allel / numpy / scipy). Nothing is written back.

• Audit an existing instance to find databases that were imported badly.

• Every analysis returns a chat summary and writes full tables as CSV under ./gigwa_results/<database>/.

Table of contents

• Gigwa MCP Server

  • Table of contents

  • Overview

  • Features

  • How it works

  • Requirements

  • Installation

  • Configuration

  • Connecting from an MCP client

  • Quick start

  • Tool reference

  • Usage scenarios

  • Output files

  • Visualizing results

    • PCA: pca_coords.csv

    • Population structure: structure_clusters.csv

    • Kinship: kinship_matrix.csv

    • Per-group diversity: diversity_by_group.csv

    • Core-collection coverage: core_collection.csv

    • UPGMA tree: tree.nwk

  • Performance & scaling

  • Limitations & disadvantages

  • Troubleshooting

  • DArTseq notes

    • Genomic positions (optional)

  • Project layout

  • Testing

  • License & contributing

Related MCP server: Cloud Life Sciences API MCP Server

Overview

Gigwa is a web platform for storing and querying genotyping data. Loading data into it and getting analyses out is normally manual (massaging xlsx into Gigwa's import format, clicking through the web UI, uploading .dart/.vcf, exporting VCFs, running pop-gen tools separately).

This server exposes Gigwa as a set of MCP tools. You talk to your MCP client in natural language; it picks the matching tool and fills in the arguments. There is no chat API of its own, meaning the "interface" is the tool list below plus your prompts.

The analysis tools are read-only: they extract genotypes (via async VCF export or paged BrAPI allelematrix), compute everything in Python, and write CSVs locally. They never modify the data in Gigwa.

Features

Import pipeline

QC & diversity (read-only)

Import-quality audit

How it works

MCP client (Claude Desktop / Code)
        │  natural language → tool call
        ▼
  gigwa_mcp (this server, stdio)
        │  GigwaClient: token auth, multipart upload, async progress, BrAPI v2
        ▼
     Gigwa REST API  ──►  genotypes (async VCF export  ‖  paged search/allelematrix)
        │
        ▼
  scikit-allel / numpy / scipy  →  chat summary + CSV under ./gigwa_results/<module>/

Analyses load genotypes through gigwa_mcp/analysis/genotypes.py:load_genotypes, which has two backends:

• method="vcf" (default) : exports the whole variant set once via async VCF and caches it on disk for reuse. Best for small/medium sets and when you will run several tools on the same run.

• method="allelematrix" : pages the genotype matrix via BrAPI search/allelematrix, honouring a server-side max_markers subset and sizing pages to the server's per-response cell cap. Best for large datasets where a full export is wasteful (see Performance & scaling).

Variant sets are addressed by their BrAPI variantSetDbId, of the form MODULE§projectNumber§run (e.g. MyDatabase§1§run1). list_content shows them.

Requirements

• Python ≥ 3.10

• A reachable Gigwa server (local or remote) and credentials.

• Optional: the minimap2 CLI on PATH for DArTseq genome-anchoring of very large genomes (otherwise the in-process mappy binding is used).

• Optional: the [viz] extra (matplotlib) to run the plotting recipes / regenerate the example figures.

Core Python dependencies (installed automatically): mcp, httpx, pandas, openpyxl, numpy, python-dotenv, scikit-allel, scipy, mappy.

Installation

From PyPI (recommended):

pip install gigwa-mcp                # core + analysis (scikit-allel/scipy)
pip install "gigwa-mcp[viz]"         # + matplotlib, for the plotting recipes

Or run it without installing into your environment using pipx or uv — handy as the command in an MCP client config (see below):

pipx install gigwa-mcp        # then: gigwa-mcp
uvx gigwa-mcp                 # run on demand, no install step

From source (for development or an unreleased version):

git clone <your-fork-url> gigwa-mcp && cd gigwa-mcp
python -m venv venv && source venv/bin/activate
pip install -e .            # core + analysis (scikit-allel/scipy)
pip install -e ".[dev]"     # + pytest, to run the test suite
pip install -e ".[viz]"     # + matplotlib, for plotting recipes / example figures

Run the stdio server directly to smoke-test:

python -m gigwa_mcp         # or: gigwa-mcp

(Normally you don't run it by hand as your MCP client launches it; see below.)

Configuration

Connection settings come from the environment, optionally seeded from a .env file in the working directory or any parent (cp .env.example .env and edit):

GIGWA_URL=http://localhost:8080/gigwa
GIGWA_USER=your_user
GIGWA_PASS=your_password
# GIGWA_TIMEOUT=120   # optional, seconds

GIGWA_URL is the Gigwa base URL without the /rest suffix (it is appended automatically). The target Gigwa may be local or remote. .env files are gitignored; keep credentials out of version control.

Connecting from an MCP client

Add a stdio server entry (Claude Desktop claude_desktop_config.json or Claude Code MCP settings). If you pip installed into a venv, point command at that venv's gigwa-mcp; with uv you can have it fetch and run the published package on demand with no separate install:

{
  "mcpServers": {
    "gigwa": {
      "command": "uvx",
      "args": ["gigwa-mcp"],
      "env": {
        "GIGWA_URL": "http://localhost:8080/gigwa",
        "GIGWA_USER": "your_user",
        "GIGWA_PASS": "your_password"
      }
    }
  }
}

Or with an explicit interpreter path ("command": "/abs/path/to/venv/bin/gigwa-mcp", no args) if you installed it into a virtual environment.

Credentials live in this config, so there is no per-chat "connect" step and every tool call authenticates on its own (token generated and refreshed automatically). To drive several Gigwa servers, register one entry each (e.g. gigwa-local, gigwa-remote) with its own GIGWA_URL/credentials and name the one you mean in the prompt.

Quick start

You talk to your MCP client in plain language; it calls the matching tool and fills in arguments (paths, thresholds, module names) from what you say. A typical first session:

More example prompts:

Tool reference

All variant-set tools take variant_set_db_id (MODULE§projectNumber§run). QC/diversity tools also accept output_dir (defaults to ./gigwa_results/<module>/) and the scaling args max_markers / method ("vcf" | "allelematrix"); see Performance & scaling.

Connection & import

QC & diversity (output files listed in Output files)

Audit

Usage scenarios

A. Import a DArTseq report, genome-anchored. Map the tag sequences once, inspect, then import reusing the positions:

"Where do these DArT markers sit on the X genome at reference.sr.mmi?" → map_dartseq_to_reference "Looks good, import report_snps.xlsx into MYDB reusing that mapping." → import_dartseq(..., positions_csv=...)

B. Vet an instance you inherited. Before trusting any analysis, triage every run for encoding artifacts:

"Scan my whole Gigwa for databases that were imported badly." → audit_import_quality Runs are ranked BROKEN / SUSPECT / OK with reasons, and the full table lands in import_quality_scan.csv.

C. Genebank cleaning. Classic data-cleaning sweep on one run:

"Check call rates, flag heterozygosity outliers, and find duplicate accessions in MYDB§1§run1." → qc_call_rate → qc_heterozygosity → qc_duplicate_accessions.

D. Diversity & structure study.

"Give me a diversity summary, a PCA, the number of clusters, and a UPGMA tree for MYDB§1§run1." → diversity_summary → diversity_pca → diversity_structure → diversity_tree.

E. Build a core collection.

"Pick a core of ~10% of accessions that captures the most allelic diversity." → diversity_core_collection(fraction=0.1).

F. Population comparisons from metadata. Provide a metadata TSV with a grouping column (e.g. country, population):

"Using meta.tsv grouped by population, compare per-group diversity and compute pairwise Fst." → diversity_by_group(metadata_tsv="meta.tsv", group_column="population") → diversity_fst(...).

Output files

Each analysis writes one or more CSVs (Newick for the tree) under ./gigwa_results/<module>/ (the audit writes to ./gigwa_results/):

Visualizing results

The tools output tables, not images, which keeps them composable. The figures below were produced from a synthetic dataset by docs/make_example_figures.py (run pip install -e ".[viz]" && python docs/make_example_figures.py to regenerate). The same recipes work on the real CSVs the tools write.

PCA: pca_coords.csv

import pandas as pd, matplotlib.pyplot as plt
df = pd.read_csv("gigwa_results/MYDB/pca_coords.csv")
groups = df["group"] if "group" in df else pd.Series("all", index=df.index)
for g, sub in df.groupby(groups):
    plt.scatter(sub.PC1, sub.PC2, s=20, label=g)
plt.xlabel("PC1"); plt.ylabel("PC2"); plt.legend(); plt.savefig("pca.png")

Population structure: structure_clusters.csv

df = pd.read_csv("gigwa_results/MYDB/structure_clusters.csv")
plt.scatter(df.PC1, df.PC2, c=df.cluster, cmap="tab10", s=20)
plt.xlabel("PC1"); plt.ylabel("PC2"); plt.title("K-means clusters"); plt.savefig("structure.png")

Kinship: kinship_matrix.csv

g = pd.read_csv("gigwa_results/MYDB/kinship_matrix.csv", index_col=0)
plt.imshow(g.values, cmap="viridis"); plt.colorbar(label="relatedness"); plt.savefig("kinship.png")

Per-group diversity: diversity_by_group.csv

d = pd.read_csv("gigwa_results/MYDB/diversity_by_group.csv").set_index("group")
d[["he", "ho", "allelic_richness"]].plot.bar(); plt.tight_layout(); plt.savefig("by_group.png")

Core-collection coverage: core_collection.csv

c = pd.read_csv("gigwa_results/MYDB/core_collection.csv")
plt.plot(c["rank"], c["coverage_fraction"] * 100)
plt.xlabel("core size"); plt.ylabel("% alleles captured"); plt.savefig("core.png")

UPGMA tree: tree.nwk

tree.nwk is standard Newick; open it directly in FigTree or iTOL, or render in Python:

from Bio import Phylo            # pip install biopython
Phylo.draw(Phylo.read("gigwa_results/MYDB/tree.nwk", "newick"))

Performance & scaling

• Small/medium runs: the default method="vcf" exports once and caches; running several tools on the same run reuses the cached genotypes.

• Large runs (hundreds of thousands of markers): pass method="allelematrix" with a max_markers cap (e.g. 2000-20000) so genotypes are sampled server-side instead of exporting a multi-GB VCF. Statistics are estimated from the sample.

• Many samples (thousands): the server caps each allelematrix response at ~10,000 cells, so at N samples a response holds ~`10000/Nmarkers, i.e. requests scale withmax_markers. Keep max_markers` modest on high-sample-count sets.

• O(samples²) tools: diversity_kinship, qc_duplicate_accessions, and diversity_tree build a samples × samples matrix (and the kinship CSV is written in full). Subsample markers and expect large output / slower runs beyond a few thousand accessions.

• The audit_import_quality tool is bounded by max_markers × max_samples per run, so it is cheap and roughly constant-cost even across a whole production instance.

Limitations & disadvantages

• Read-only analysis. QC/diversity/audit never write results back to Gigwa; you get CSVs locally. (Import tools do write to Gigwa.)

• No built-in plotting. Tools emit CSV/Newick; use the recipes above (matplotlib/Bio.Phylo) to make figures.

• allelematrix has no session cache. Unlike the VCF backend, the allelematrix path re-fetches genotypes on each tool call, so running many tools on the same large set re-downloads the sample each time.

• diversity_structure is a lightweight heuristic. It is PCA + K-means with a pseudo-F (Calinski-Harabasz) K suggestion; there is no true admixture model. On weakly or continuously structured data pseudo-F tends toward k_max; the per-K table is the real output and the tool warns when clustering is degenerate. For formal ancestry use a dedicated tool (ADMIXTURE / sNMF) on an exported VCF.

• Diploid-biallelic assumptions in places (IBS dosage 0/1/2, collapsed-token decode).

• Grouping uses a metadata TSV, not server attributes. Some Gigwa builds do not expose BrAPI germplasm/sample/attribute endpoints, so diversity_fst / diversity_by_group take groups from groups_json or a metadata TSV rather than querying Gigwa.

• VCF export downloads the whole variant set regardless of max_markers; use method="allelematrix" to subsample large sets.

• Genome anchoring needs minimap2 + a reference, and streaming very large indexes is I/O-bound.

• Single-threaded Python compute; large matrices are held in RAM.

Troubleshooting

• Auth / "Missing required environment variable(s)". Ensure GIGWA_URL, GIGWA_USER, GIGWA_PASS are set (env or .env). GIGWA_URL must omit the /rest suffix.

• VCF import rejected / "not bgzipped". Gigwa needs BGZF, not plain gzip. Recompress: gunzip -c f.vcf.gz | bgzip > f.bgz.vcf.gz (htslib bgzip).

• Implausible ~95% heterozygosity after a DArT import. That is Gigwa's built-in DArT parser mis-calling the 2-row format. Use import_dartseq (it calls genotypes in Python and imports a standard VCF) instead of importing the raw DArT report (see below).

• diversity_fst / diversity_by_group report "no groups matched". Check that id_column values in your TSV match the accession names (or callset ids) in the run.

• Large set feels slow. Use method="allelematrix" + a smaller max_markers, and avoid the O(samples²) tools on many thousands of accessions.

DArTseq notes

DArTseq SNP reports use the classic 2-rows-per-marker layout (a reference-allele row and a SNP-allele row, each cell 1/0/-); Silico-DArT reports are 1 row per clone (dominant presence/absence). import_dartseq does the genotype calling in Python and emits a standard VCF, imported through Gigwa's verified VCF path:

(ref=1, alt=0) -> 0/0   (ref=0, alt=1) -> 1/1
(ref=1, alt=1) -> 0/1   otherwise      -> ./.   (missing / no allele detected)

This deliberately bypasses Gigwa's built-in DArT parser, which might mis-call the 2-row format (there are cases that it imports reference homozygotes as heterozygous, producing implausible ~95% heterozygosity). SNP and Silico use different allele models; import them as separate runs unless you specifically intend to combine them.

Genomic positions (optional)

DArTseq markers have no genomic coordinates, so by default they are placed on a single Unmapped contig at sequential positions. If you have a reference genome FASTA, the marker tag sequences (AlleleSequence, ~69 bp) can be aligned to it with minimap2 to infer real chromosome/position/strand:

• map_dartseq_to_reference(snp_xlsx, reference_fasta) → a dartseq_positions.csv report (uniquely mapped / multi / unmapped), for inspection.

• import_dartseq(..., reference_fasta=...) → imports uniquely-mapped markers genome-anchored (minus-strand alleles complemented, output coordinate-sorted, one marker per genomic site); unmapped markers stay on Unmapped.

• import_dartseq(..., positions_csv=...) → reuse a dartseq_positions.csv from a previous run instead of re-aligning. Recommended for large genomes: align once, inspect, then import without paying the alignment cost again.

reference_fasta may be a FASTA (.fa/.fa.gz) or a prebuilt minimap2 .mmi index. By default the minimap2 CLI backend is used when available: it streams over multi-part indexes with bounded RAM, so very large (multi-gigabase) genomes work on modest machines. The in-process mappy backend (backend="mappy") loads the whole index into RAM instead.

Prebuild an index once (tuned for the ~69 bp tags) and reuse it:

minimap2 -x sr -d reference.sr.mmi reference.fasta   # build once
# then pass reference.sr.mmi as reference_fasta

Project layout

gigwa_mcp/
  __main__.py           # python -m gigwa_mcp → stdio server
  config.py             # .env / env loading (GIGWA_URL/USER/PASS/TIMEOUT)
  client.py             # GigwaClient: auth, multipart upload, progress, BrAPI calls
  server.py             # FastMCP instance + get_client()
  importers/
    dartseq.py          # DArTseq xlsx → standard VCF (2-row genotype calling)
    refmap.py           # minimap2 tag → reference mapping
  analysis/
    genotypes.py        # load_genotypes (VCF / allelematrix backends), GenotypeMatrix
    stats.py            # pure pop-gen stats (MAF, He, PIC, IBS, GRM, allelic richness …)
    genebank.py         # core-collection + UPGMA helpers
    results.py          # output-dir resolution + CSV writing
  tools/                # @mcp.tool() wrappers: connection, genotype, metadata, qc,
                        #   diversity, audit
scripts/                # run_import_audit.py, run_qc_diversity_validation.py (generic)
docs/                   # make_example_figures.py + img/ (README figures)
tests/                  # pytest suite (mocked client + synthetic fixtures)

Testing

pip install -e ".[dev]"
pytest

test_client.py covers auth/token-refresh, multipart assembly and progress polling with a mocked transport; test_dartseq_convert.py checks the conversion against synthetic SNP/Silico fixtures; test_stats.py / test_genebank.py verify the pop-gen and genebank statistics against hand-computed values; test_genotypes.py exercises VCF parsing + callset-name mapping with a mock client. The suite needs no live Gigwa server.

License & contributing

Released under the Apache License 2.0 © 2026 Anestis Gkanogiannis anestis@gkanogiannis.com (see also NOTICE).

Issues and pull requests are welcome. Please run pytest before submitting, keep new analysis logic in pure, unit-tested helpers under gigwa_mcp/analysis/, and avoid committing data, credentials, or result files (these are gitignored).