aipoch-openscience-mcp

One-install access to 23 life-sciences MCP servers — PubChem, ChEMBL, Ensembl, UniProt, PDB, AlphaFold, GEO, ArrayExpress, gnomAD, ClinVar, GWAS Catalog, GTEx, OpenAlex, PubMed, ClinicalTrials.gov, and more.

The servers are vendored from Claude Science and fully self-contained: the plugin bootstraps its own Python virtualenv on first run, so no Claude Science installation is required.

What you get

A single plugin, openscience, that registers 23 MCP servers (233 tools total). Each server maps to a public life-sciences database group:

Server	Sources	Server	Sources
chemistry	PubChem, ChEBI, Rhea, BindingDB	genomes	Ensembl (incl. VEP), UCSC
chembl	ChEMBL	genes-ontologies	MyGene, UniProt, GO, Reactome, OLS
zinc	ZINC	variants	gnomAD, ClinVar, dbSNP
pubmed	PubMed / NCBI E-utilities	human-genetics	GWAS Catalog, eQTL, FinnGen
literature	OpenAlex, arXiv	clinical-genomics	ClinGen, CIViC, Open Targets
biorxiv	bioRxiv	expression	GTEx
clinical-trials	ClinicalTrials.gov	regulation	ENCODE, JASPAR, UniBind
drug-regulatory	openFDA	protein-annotation	InterPro, Pfam, HPA, STRING
structures-interactions	PDB, AlphaFold, EMDB, Complex Portal, IntAct	rna	Rfam
omics-archives	GEO, ArrayExpress, PRIDE, MGnify, MetaboLights	cancer-models	cBioPortal
cellguide	CELLxGENE cell types	biomart	BioMart
research-resources	Grants.gov, Antibody Registry

All tools are read-only retrieval against public databases.

Requirements

• A host with MCP support — Claude Code (plugin support) or WorkBuddy (skill install, see below).
• Python ≥ 3.11 on your PATH (used once to build the plugin's private virtualenv).
• Network access on first run (to pip install the pinned dependencies).

Install

Installation depends on your agent host. Currently supported:

Claude Code

/plugin marketplace add aipoch/openscience-mcp
/plugin install openscience@aipoch-openscience

Restart Claude Code when prompted. On the first tool call, the plugin builds its virtualenv and installs dependencies (about a minute, one time). Subsequent starts are instant.

WorkBuddy

This repo ships a self-contained skill, openscience-mcp-connector, under .workbuddy/skills/openscience-mcp-connector.

python .workbuddy/skills/openscience-mcp-connector/scripts/install.py \
  --repo "$(pwd)" --data-dir ~/.cache/openscience-mcp
# then enable the new connectors in WorkBuddy's connector settings

Pass --servers chemistry,pubmed to install a subset and avoid loading all 233 tools at once. See the skill's SKILL.md for the full workflow.

More agent frameworks and MCP hosts are on the roadmap.

Configuration (optional)

When you enable the plugin, Claude Code prompts for two optional values:

• Contact email (NCBI etiquette) — NCBI E-utilities (PubMed, dbSNP, ClinVar, GEO) ask callers to identify themselves. A few tools (e.g. pubmed metadata) require it; set this to use them.
• NCBI API key — optional, raises the NCBI rate limit from 3 to 10 requests/second.

Both can be left blank; servers that don't need them work regardless.

Usage

Just describe what you need, or name a source. Examples:

• "Look up the PubChem properties of aspirin." → chemistry
• "Find GWAS associations for height." → human-genetics
• "Get the AlphaFold structure for UniProt P04637." → structures-interactions
• "Search OpenAlex for CRISPR review authors." → literature

Verification

Every server is covered by an integration test harness. The launcher's own unit + integration tests run standalone:

bash tests/launch_smoke.sh   # unit + integration tests for the plugin launcher

Per-server results — every server booted, all 233 tools enumerated, and a representative live call per server — come from the integration harness (needs network; upstream databases may rate-limit):

python tests/test_all_servers.py   # boots each server, enumerates tools, runs representative calls

How it works

• plugins/openscience/runtime/lib/ — the vendored MCP server packages (Python).
• plugins/openscience/bin/launch.sh — bootstraps a private virtualenv in the plugin's data directory and launches a server over stdio. No Claude Science dependency.
• plugins/openscience/.mcp.json — registers the 23 servers with Claude Code.

Roadmap

Phase 1 (this release) ships the 23 servers over stdio. Phase 2 introduces a gateway with dynamic tool discovery to cut process count and context cost. See ROADMAP.md.

License

The vendored server code originates from the Claude Science bundled bio-tools MCP servers; see plugins/openscience/runtime/VENDOR_SOURCE.txt for provenance. Each upstream database has its own terms of use — review the individual source licenses for your use case.