Data sources.
Thirteen public biomedical databases, reached directly. Essarion Science pulls records — sequences, structures, expression series, trial registrations, variant calls — not summaries of records.
§ 01The roster
Each connector is a first-class citizen of the tool registry, individually switchable per run from the settings panel.
| Connector | Provides |
|---|---|
| ncbi | PubMed biomedical literature via NCBI E-utilities. |
| europepmc | Life-science literature plus bioRxiv and medRxiv preprints. |
| openalex | Open scholarly index across all disciplines. |
| uniprot | Protein sequence and annotation records from UniProtKB. |
| ensembl | Gene annotations — location, biotype, cross-references. |
| rcsb_pdb | Experimental 3D structures and coordinate files from the PDB. |
| alphafold | Predicted protein structures and coordinate URLs, by UniProt id. |
| chembl | Bioactive molecules, SMILES, and assay phases. |
| reactome | Curated biological pathways for a gene, molecule, or process. |
| clinvar | Clinical significance of germline and somatic variants. |
| geo | Public gene-expression datasets and series, bulk and single-cell. |
| opentargets | Target ↔ disease association scores, for target ranking. |
| clinicaltrials | Registered clinical trials — started and completed — by topic. |
Returns the live catalog — every connector and skill with its id, label, category, description, and default_on flag. The settings UI is rendered from exactly this payload, so it never drifts from what the engine actually has.
§ 02Key-free by design
Every connector and skill talks to a public scientific API or runs local compute. There is nothing to configure to use them and no credential to obtain. Optional keys exist only to raise rate limits or unlock a specific capability:
| Variable | Effect |
|---|---|
| NCBI_API_KEY | Raises PubMed limits to 10 requests/second, up from 3 anonymous. |
| SEMANTIC_SCHOLAR_API_KEY | Raises Semantic Scholar Graph API limits. |
| OPENALEX_MAILTO | Contact email for the OpenAlex polite pool. |
| CROSSREF_MAILTO | Contact email for the CrossRef polite pool. |
| BIO_HUB_API_KEY | Protein design via ESM3 inverse folding, and a reliable folding fallback. Folding still works key-free via ESMFold. |
§ 03Choosing your sources
Connectors resolve per run from the settings map sent with the request, so the same account can run a narrow study and a broad one back to back. Two habits are worth forming:
- Narrow when the question is narrow. A pure structural-biology question does not benefit from trial registrations; turning off what cannot contribute keeps the loop's attention on what can.
- Leave literature on. Effect sizes come from the literature pool, and the evidence gates in Verification count independent sources and distinct labs. Disabling literature connectors is the fastest way to make a run unable to emit a paper.
§ 04How literature is handled
Literature retrieval runs in two stages. A search stage pulls candidates up to the depth mode's target — 16, 36, or 72 — and a deep-read stage takes the top slice of those (6, 18, or 32) into the effect-size extractor, which is where pooled statistics come from.
On exhaustive runs, PDF vision is enabled: figures are extracted from source PDFs rather than described from their captions.
§ 05Primary data
Beyond the biomedical connectors, runs can draw on primary datasets for context and triangulation. The available set widens with depth: scan reaches the World Bank and Our World in Data; deep_dive adds FRED, SEC, OECD, and trial registrations; exhaustive adds IMF, BLS, Eurostat, and GDELT.
Primary-data correlations are one of the inputs to cross-source triangulation — a pattern that appears in both the pooled literature and an independent dataset is qualitatively stronger evidence than the same number repeated from one source.
§ 06Provenance
Every retrieved record is kept with the connector that produced it and the query that found it, and every notebook cell carries provenance for the data it consumed. References in a finished paper are real database records, so a reader can resolve any one of them back to its source and check it. That is the whole point: an answer you cannot check is one you take on faith.