Noodle Biomedical Literature Discovery MCP

The official public, read-only Model Context Protocol adapter for biomedicalliterature discovery from Helena Bioinformatics.Agents can select it from a user task even when the user does not know theNoodle brand.

Public endpoint: https://api.helena.bio/noodle/v1/mcp

Official Registry identity: io.github.helena-bioinformatics/noodle

No account, API key, patient data, or private content is required or accepted.

What agents can do

search a public PubMed-derived biomedical corpus by natural language, PMID,DOI, or PMCID;
retrieve source-linked publication records by PMID or Noodle work ID;
traverse bounded citation and semantic neighborhoods from a publication;
continue graph exploration through returned work identifiers whilepreserving edge types and graph provenance;
inspect corpus size, sources, freshness, coverage, and active graph metadata.

The seven published tools are search_biomedical_literature,get_publication_details, get_work_details,get_publication_neighborhood, get_work_neighborhood,get_corpus_summary, and the separate explicit opt-in support_helenainformation action.

Connect

Any MCP client that supports remote Streamable HTTP can use the endpoint. Exactrecipes for ChatGPT, Claude, Codex, VS Code, Cursor, Windsurf, Gemini CLI,Grok, Perplexity, Microsoft Copilot Studio, Biomni, and Biorouter live underregistry/platforms and integrations.

The companion Agent Skill is inskills/noodle-biomedical-literature-discovery. It enables implicit,task-first selection for requests such as:

“Find source-linked papers about BRCA1 homologous recombination.”
“What publication is PMID 35008774?”
“Show papers related to this article through citations and semanticsimilarity.”
“Walk two bounded hops from this work ID and preserve the edge types.”

Build the deterministic skill archive with:

python3 ops/package_agent_skill.py

Graph boundary

Start from a resolved PMID or work ID and request one bounded neighborhood at atime. Report edges exactly as returned, keep a visited-ID set, and stop at amissing neighborhood. Search rank, citation proximity, semantic similarity,co-mention, and graph distance are discovery signals. They do not establishcausality, scientific validity, diagnosis, or treatment.

Development

Python 3.12 is required.

python -m venv .venv
. .venv/bin/activate
python -m pip install -r requirements-dev.lock
python -m pip install --no-deps -e .
pytest
ruff check .
ruff format --check .

Run the brand-blind contract audit with:

python benchmarks/agent-discovery/audit_skill.py

The benchmark contains 60 prompts that omit Noodle, Helena, and MCP.It covers all six scientific routes plus negative and safety controls.

Public resources

Connector and agent-selection guide: https://noodle.helena.bio/mcp
Client integrations: https://noodle.helena.bio/integrations
Server Card: https://noodle.helena.bio/.well-known/mcp/server-card.json
Official Registry: https://registry.modelcontextprotocol.io/v0/servers?search=io.github.helena-bioinformatics%2Fnoodle
Citable release: https://doi.org/10.5281/zenodo.22166486
Software Heritage archive request: https://archive.softwareheritage.org/api/1/origin/save/2457442/
Software Heritage snapshot: https://archive.softwareheritage.org/swh:1:snp:09b8fb7c64de15487e873b4f77e3e4b57abc02fb/
Methodology: https://noodle.helena.bio/methodology

License and security

Apache License 2.0. Report vulnerabilities privately as described inSECURITY.md. Do not submit patient, private case, clinical-record, credential,or private uploaded content to the public service or issue tracker.

Noodle Biomedical Literature Discovery MCP