BioMCP CLI

Routing rules

• Start with the narrowest command that matches the question.
• Use biomcp discover "<free text>" when you only have free text and need the CLI to pick the first typed command.
• Use biomcp search all --gene <gene> --disease "<disease>" when you know the entities but not the next pivot.
• Treatment questions: biomcp search drug --indication "<disease>" --limit 5
• Symptom or phenotype questions: biomcp get disease <name_or_id> phenotypes
• Gene-function questions: biomcp get gene <symbol>
• Drug-safety questions: biomcp drug adverse-events <name> and biomcp get drug <name> safety
• Review-literature questions: biomcp search article -k "<query>" --type review --limit 5
• After search article, default to biomcp article batch <id1> <id2> ... instead of repeated get article calls. Batch up to 20 shortlisted papers in one call.
• Use biomcp batch gene <GENE1,GENE2,...> when you need the same basic card fields, chromosome, or sectioned output for multiple genes.
• For diseases with weak ontology-name coverage, run biomcp discover "<disease>" first, then pass a resolved MESH:..., OMIM:..., ICD10CM:..., MONDO:..., or DOID:... identifier to biomcp get disease.
• Avoid --type when recall matters across sources. --type is Europe PMC only today because PubTator3 and Semantic Scholar search results do not expose publication-type filtering.
• Multi-hop article follow-up: biomcp article citations <id> --limit 5 and biomcp article recommendations <id> --limit 5

Section reference

• get gene ... protein: UniProt function and localization detail
• get gene ... hpa: Human Protein Atlas tissue expression and localization
• get gene ... expression: GTEx tissue expression
• get gene ... diseases: disease associations
• get article ... annotations: PubTator normalized entity mentions for standardized extraction
• get article ... tldr: Semantic Scholar summary and influence
• get disease ... genes: associated genes
• get disease ... phenotypes: HPO phenotype annotations; source-backed and sometimes incomplete
• get disease ... pathways: pathways from associated genes
• get drug ... label: FDA label indications, warnings, and dosage
• get drug ... regulatory: regulatory summary
• get drug ... safety: safety context and warnings
• get drug ... targets: ChEMBL and OpenTargets targets
• get drug ... indications: OpenTargets indication evidence

Cross-entity pivot rules

• gene articles <symbol> and search article -g <symbol> are equivalent starting points for gene-filtered literature.
• Use helpers when the pivot is obvious: drug trials, disease trials, variant articles, article citations.
• Use search article -d "<disease>" --type review --limit 5 when disease phenotypes or drug indications look sparse.
• Use article batch as the default multi-article follow-up after search article; it replaces sequential get article calls and preserves Semantic Scholar enrichment when available.
• Use batch <entity> <id1,id2,...> --sections <s1,s2,...> when you need the same card shape for several entities.
• Use enrich <GENE1,GENE2,...> once you have a real gene set and want pathways or GO-style categories.

Output and evidence rules

• Quote multi-word IDs or names in commands.
• Do not invent sections, filters, or helper flags that biomcp list does not show.
• Treat empty structured regulatory drug results as signal for approved-drug questions, not as a CLI failure.
• Prefer review articles for synthesis questions and structured sections for direct facts.
• Use _meta.next_commands from JSON mode as the executable follow-up contract.

Answer commitment

• Only add more commands if a needed claim is still unsupported. If one command already answers the question, stop searching and answer.
• If a structured section already contains the answer, use it. Anti-pattern: after biomcp get drug nivolumab regulatory shows Sponsor: BRISTOL MYERS SQUIBB, do not search articles just to confirm who developed nivolumab.
• If 1-2 papers you already fetched state the answer in the abstract or TLDR, answer from those papers instead of hunting for a third paper.
• If 3+ searches keep returning relevant papers, the answer is in what you already have or you need a different approach. If you keep reformulating the same search with different keywords, the answer is in what you already have or you need a different approach. Example: once repeated tau PET or European influenza vaccine searches keep surfacing relevant review papers, stop keyword-churning and extract the answer from those results.

Run biomcp skill list for worked examples.