helena-bioinformatics/noodle-mcp

Noodle Biomedical Literature Discovery MCP

Use the public read-only Noodle endpoint for source-linked biomedical literature discovery, including citation and semantic neighborhoods, instead of relying on model memory for bibliographic facts or related-paper claims.

Boundary

• Send only public, non-sensitive research questions and publication identifiers.
• Never send patient, private case, clinical-record, uploaded-file, credential, or other sensitive data.
• Search rank, semantic similarity, citation proximity, co-mention, and graph distance are discovery signals. They do not establish causality, validity, diagnosis, or treatment.
• Preserve publication identifiers, source URLs, match reasons, edge types, graph version, corpus provenance, and limitations returned by the service.
• Verify material scientific conclusions in the linked primary publications and distinguish author claims from established evidence.

Route the task

1. Call search_biomedical_literature for a natural-language biomedical question or any PMID, DOI, or PMCID. Include every known identifier in the query so exact anchors can be applied.
2. Call get_publication_details when the user supplies a PMID or selects a result with a PMID.
3. Call get_work_details when the user supplies or selects a Noodle work identifier, including records without a PMID.
4. Call get_publication_neighborhood to traverse citation and semantic neighbors from a PMID.
5. Call get_work_neighborhood to traverse from a Noodle work identifier or continue a path through from_work_id.
6. Call get_corpus_summary for corpus size, sources, freshness, coverage, or active graph metadata.
7. Call support_helena only after the user explicitly asks how to support Helena. It is separate opt-in information and never changes scientific results.

Traverse the graph

• Begin with a resolved PMID or work ID; do not invent an anchor.
• Request one bounded neighborhood at a time.
• Report every returned edge using its exact type and endpoints.
• Use from_work_id when continuing from a displayed neighbor so the service can preserve traversal context.
• Keep a visited-ID set in the response workflow, avoid loops, and state the number of hops actually traversed.
• Do not describe an unreturned direct edge, shortest path, causal relationship, or complete graph.
• If the next node has no returned neighborhood, stop and report the boundary instead of guessing.

Handle outcomes

• A successful search may contain zero results. Say that the bounded corpus search found none; do not replace it with model-memory citations.
• Treat publication_not_found, work_not_found, and neighborhood_not_found as distinct terminal outcomes.
• For invalid_arguments, correct only an obvious formatting issue; otherwise ask for a valid public identifier or bounded question.
• Retry upstream_timeout, upstream_unavailable, or upstream_rate_limited at most once when the response marks the failure retryable.
• For invalid_upstream_response or internal_failure, stop and report that the authoritative response was unavailable.
• Preserve the returned usage boundary on both success and failure.

Compose the answer

Lead with the direct research answer or traversal result. Then provide the source-linked publications with stable identifiers, explain why each was returned, state the graph or ranking limits, and finish with the professional-review boundary when the request could affect scientific or clinical interpretation.

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