Abstract
TogoMCP is a Model Context Protocol (MCP) server that lets large language model (LLM) agents querylife-science knowledge graphs in SPARQL, guided by per-database schema documents called MIE(Metadata Interoperability Exchange) files. During the DBCLS BioHackathon 2026 (BH26), the TogoMCPgroup set out to extend the server, refine the MIE files, add databases, learn what makes a goodSPARQL example, and turn use cases into reusable agent skills. In the first six days of the eventwe published eleven releases (v2.12.2 to v2.20.0) and grew the catalogue from 37 to 45 databases:Fanta.bio, WikiPathways, IDSM, PubCaseFinder, LIPID MAPS, SwissLipids and MarpolBase, plus BH26Microbes, an experimental dataset of KEGG Orthology assignments for 57.6 million prokaryoticproteins built for the BioHackathon. Five of them are served from endpoints outside RDF Portal,and a sixth from a QLever engine inside it rather than Virtuoso, and they broke assumptions therest of the corpus had taught, such as named-graph pinning, federation with `SERVICE`, literaltyping, and even that a result set is complete. For the first time, MIE files were written bypeople other than the server's maintainer, one of them by the maintainer of the database itdescribes, and a second group member built TogoCX, a companion MCP server that returns databaseedges with instructions on how their claims may be stated. We made every worked example in the MIEcorpus (now 420) assert its recorded result against the live endpoint in continuous integration.The first full comparison found drift that execution-only checks had missed, most seriously aNANDO release that made an example join silently miss about 88% of mapped diseases. Onboarding thenew databases uncovered a series of quantified "silent wrong answers", queries that returnplausible results rather than errors, including a category count inflated 3.1-fold by an upstreamdata defect and a gene enumeration that returns 10,000 of 18,080 rows without an error or awarning. We also added two PubCaseFinder tools for phenotype-driven rare-disease diagnosissupport, began serving analysis workflows (agent skills) from the server, fixed failure modesfound in production call logs, made TogoID errors suggest working conversion routes, and prepareda LOTUS natural-products graph for hosting on RDF Portal. We argue that for LLM-facing schemadocumentation, "the example still runs" is not evidence that it is still correct.