methods-mcp

Bioconductor method adapters for the md-lab substrate. Each adapter wraps one method (limma, fgsea, DESeq2, …) as an MCP tool backed by a small FastAPI-in-Docker container that runs R via rpy2.

Replaces the originally planned playbase-mcp. OmicsPlayground is CC BY-NC-ND (commercial-unsafe); we wrap Bioconductor directly. See md-lab/NOTES.md §3.

Status

v0.1.0 skeleton. Only limma is scaffolded, and its container doesn't exist yet — tool calls return status: "container_unavailable" until the container ships. This repo exists to prove the adapter pattern end-to-end.

Architecture

Cowork / Codex / Jupyter
    ↓ MCP tool call: limma(intensity_parquet=..., contrast=..., ...)
methods-mcp (FastMCP, this repo)
    ↓ HTTP POST /run
method-limma container (FastAPI + rpy2, separate repo)
    ↓ rpy2
R: limma::eBayes(limma::lmFit(...))
    ↑ result parquet
methods-mcp
    ↑ {results_parquet, summary, methods_sentence, provenance}
Caller

Each adapter has four files: manifest.yaml, adapter.py, Dockerfile, README.md. See src/methods_mcp/adapters/limma/ for the reference shape.

Core ship list (W2–W3)

Single-cell (Seurat, monocle3, Azimuth) is also in v0.1 scope per interview answers §Q4.

Install

pip install -e ".[dev]"
pytest
# run server
python mcp_server.py            # stdio transport
python mcp_server_http.py       # HTTP transport, port 7102

Community registry

Once the pattern holds, any contributor can ship an adapter package: manifest.yaml + adapter.py + Dockerfile + README.md. md-lab methods install <name> fetches from the central YAML index in webwebb56/md-methods-registry. Homebrew-style install-by-fetch, not npm-style auto-resolution.