FAQ

The questions partners and users ask most. Each answer links the full documentation.

Is Chinese supported?

Skills are authored in English (ecosystem convention for stable cross-harness triggering); the agent replies in the prompt's language — a Chinese prompt yields a Chinese report, with a true-run demo as proof. The WorkBuddy connector ships official zh descriptions and example prompts. See the zh demo

Do I need API keys?

No — every demo on this site ran keyless. The only exceptions: OncoKB annotations require a token, and DisGeNET falls back to OpenTargets without one. Keys raise rate limits (e.g. NCBI 3→10 req/s). MCP doc §5.3

Does it work from mainland China?

Yes — onboarding/uv skills and the WorkBuddy connector use npmmirror/China mirrors automatically; upstream biomedical APIs may need a proxy (HTTP(S)_PROXY is honored).

Why do the demo reports open directly here, when GitHub can't render them?

Because this site embeds the byte-exact final_report.html committed in the repository — the same file whose provenance is pinned by the repo commit. Nothing is re-rendered for display. Demos

Can the agent fabricate references?

No — and not merely by instruction. Since deep-research 1.5+ the citation pipeline is key-based end-to-end: workers append every citable source to an evidence ledger, the draft is authored with [@pmid:...] cite-key markers, and the render script is the single numbering authority (it hard-fails on any key without a ledger record); an independent auditor then re-checks the rendered report (contiguous [1]..[N], NCBI cross-check). Hand-assembled bibliographies are structurally impossible.

How many tokens / how long do runs take?

Measured (opencode 1.18.30, deep-research 1.7.0 with parallel dispatch): interview 33 s; PubMed parse 56 s; deep research EN 545 s / ZH 649 s; structural figure incl. dependency install 1064 s. The MCP probes are token-free.

What about data privacy?

All queries hit public biomedical APIs only; reports and figures stay local to your project; the repo contains no telemetry; demo artifacts contain no credentials.

License and commercial use?

Apache-2.0 (repo LICENSE and the upstream biomcp-ts project). Commercial use and derivatives permitted with the license notice retained; citation metadata in CITATION.cff.

Does it work in Cursor / Codex / WorkBuddy?

Yes — any harness that reads SKILL.md and can run an MCP server. See the compatibility matrix in the agent doc §4.5.

How do I reproduce the demos?

Clone the repo and run node demos/run-demo.mjs --list; replay any case with --only <id> --publish. Agent cases spend LLM tokens (manual-run only); the MCP probes are token-free. demos/README.md

All fourteen questions: agent doc §7