biorun

Biology’s foundation models, run exactly as published.

Single-cell and protein models behind one interface: in the browser, over an API, or from your coding agent. Every result comes with a baseline and a record of how it was made.

Get started Connect an agent →

Results you can trust

Faithful

Preprocessing matches each model’s reference code token for token, and every weight is checked as loaded. Input a model can’t use is refused, with the fix.

Honest

A simple baseline is computed beside every result, on the same data, so you see what the model adds.

Reproducible

Each output records the model, its weights, the preprocessing version and the seed. The same request gives the same answer.

Published models, ready to run

Single-cell

  • Geneformer
  • scGPT
  • SCimilarity
  • State Embedding
  • State Transition
  • TranscriptFormer
  • UCE

Protein

  • ESM-2
  • ESMC
  • Cell embeddings
  • Cell type annotation
  • Cell search in a reference atlas
  • Perturbation prediction
  • Protein embeddings
  • Variant effect scores

From your agent

Any agent that speaks MCP connects with one address and signs in through your browser. It picks a model, checks your file, runs it and brings back the result with its baseline.

In a terminal

claude mcp add --transport http \
  biorun https://app.getbiorun.com/mcp

Where it runs

On a GPU started for your run and stopped when it ends, or on your team’s own machines: a cluster or a workstation, connected with one command.

Your data

We keep no data by default. An uploaded file, and the file a run produces, are deleted after 24 hours.