Herb Tcm MCP server

@pipeworx/herb-tcm

HERB 2.0's Traditional Chinese Medicine knowledge base — herbs, ingredients, gene targets, diseases, PubMed-cited papers and GEO transcriptomic experiments — proxied live from herb.ac.cn, with every relationship row tagged an evidence_tier so a statistical prediction is never mistaken for proof of efficacy.

Part of Pipeworx — an MCP gateway connecting AI agents to 1576+ live data sources.

Tools

Chinese characters, pinyin, English/Latin names, gene names/aliases, disease names, or a HERB id itself. category is one of herb / ingredient / target / disease (default herb).

• herb_search(keyword, category?) — resolve a name to an id. Accepts

herb/ingredient/target/disease list (7,263 / 49,258 / 12,933 / 28,212 rows respectively).

• herbbrowse(category, page?, pagesize?) — page through the full

for one id: composition, traditional-use summary, and every predicted or literature-backed target/disease relationship, each carrying its evidencetier. Relationship tables are paged (fleet #1399 — an unpaged well-studied herb ran to 806 KB): each comes back as {total, offset, limit, returned, truncated, rows} where total is the TRUE upstream row count and truncated: true says more rows exist — so a 25-row slice can never be mistaken for 25 rows existing. limit defaults to 25 (the pack's precedent, set by herbpapers), max 200; offset pages; sections (e.g. ["summary","herbingredient"]) fetches composition without pulling thousands of predicted disease rows, and an unknown section name errors loudly rather than silently returning nothing. We chose client-side slicing over a sections-only design because detailapi has no server-side paging (one upstream call returns everything regardless), so the slice costs nothing extra and the envelope keeps the caller honest. ingredient_alias (a small synonym list) is never paged.

• herb_detail(id, category, limit?, offset?, sections?) — the full record

list PubMed-cited references, each tagged human_clinical or laboratory.

• herbpapers(drugtype?, experimenttype?, sortby?, limit?, offset?) —

the specific targets/diseases it reports.

• herbpaperdetail(paper_id) — one reference's bibliographic record plus

— list GEO-deposited herb/ingredient-vs-control transcriptomic experiments.

• herbexperiments(drugtype?, species?, experiment_type?, limit?, offset?)

for one experiment: top up/down genes, enriched GO/KEGG terms, connectivity-map hit counts. Always evidencetier: computationalprediction.

• herbexperimentdetail(experiment_id) — differential-expression results

Evidence tiers

Every relationship row carries one of:

herbtarget / herbdisease (statistical enrichment, p-value + FDRBH columns), ingredienttarget / ingredientdisease / targetdisease (curated cross-references with no clinical backing) and every herbexperimentdetail row are computationalprediction. drugpapertarget / drugpaperdisease (PubMed-cited) are split into humanclinical or laboratory by resolving each cited paper's own "Experiment type" — see Data sources below for how.

Auth

Shortened. The full README is on GitHub.