dsh-biomni

Hand a biomedical tool library to an agent and the sneakiest failure is not an error — it is silence. The dependency is missing, so it quietly writes its own PubMed client. The data directory does not exist, so it fabricates a path. The result arrives confident, indistinguishable from a real one. dsh-biomni enforces a single rule: never advertise what this machine cannot deliver. Every one of the 218 research functions, 76 datasets, and 113 bioinformatics CLI tools must pass an environment probe before appearing in the skill catalog. What is missing is named out loud, with the package that would fix it and the price in megabytes.

The core experience centers on a persistent Python interpreter: imports, dataframes, and fitted models survive across calls within a session, so the agent works in small steps instead of resending a full script every time. An interpreter idle for thirty minutes is retired and the agent is told its namespace is empty — rather than hitting a NameError on the next call. The 21 module skills load on demand, cutting 218 functions from roughly 26k to about 1.6k tokens of context. The Settings panel gives an honest availability report and tags the four opt-in extras that cost 494 MB for just 7 more functions as expensive, not free.

If you are a biomedical, pharmacology, or computational-biology researcher who drives analysis pipelines through DeepSeek Harness and wants the agent to actually call samtools, query_pubmed, or a molecular-docking function instead of reinventing them, dsh-biomni is the bridge that plugs Biomni into the agent — no forking, no patching, install and go.