Running reproducible bioinformatics pipelines from your AI assistant

There is a gap in the way AI assistants meet bioinformatics, and it is a specific one.

Ask a capable model to help with a genomics analysis today and two of the three hard parts are already handled. It can reach the data: there are Model Context Protocol (MCP) servers that wrap dozens of biomedical databases, so an assistant can look up a variant, a gene, a clinical trial. And it can plan the analysis: there are agent frameworks that reason about which tools to run in what order.

The third part is the one that was missing, and it is the one that actually matters if you are going to trust the answer: running the pipeline, and being able to prove what ran.

An AI that "does the analysis for you" is only useful if the result is reproducible. Otherwise it is a black box with a confident tone. In science that is worse than useless.

So I built FlowProof.

What it is #

FlowProof is an open-source MCP server. It lets an AI assistant run a real bioinformatics pipeline and hand back results whose provenance can be independently verified.

Concretely, when a pipeline runs, FlowProof emits a Workflow Run RO-Crate: a standard, machine-readable record capturing the workflow and its version, the container images and their digests, the resolved tool versions, the exact parameters, and SHA-256 checksums of every input and output. That record is what turns an AI-produced result from "trust me" into "here is the exact recipe, reproduce it byte for byte."

The tool surface an assistant sees is small and honest:

• list the available pipelines
• describe a pipeline's inputs and parameters
• run a pipeline
• check run status and fetch results (with checksums)
• get the provenance record

That last one is the whole point.

Why provenance is the hard part #

It is easy to shell out to a workflow engine and stream some logs back. It is harder to make the run trustworthy.

Reproducibility in bioinformatics fails in quiet ways: a slightly different tool version, a container that drifted, a parameter nobody wrote down. The community's own literature is full of "the code and data were available and we still could not reproduce it." The fix is discipline: pin the containers, record the versions, checksum the artifacts, and emit it in a standard format rather than a homemade one. FlowProof uses the Workflow Run RO-Crate profile precisely so the provenance is portable and not another bespoke JSON blob.

How it runs #

The execution backend is an interface. There is a deterministic mock backend for development and CI, and a Nextflow backend that shells out to a real nextflow run with a container profile. Pipelines are plugins registered by manifest, so short-read and long-read workflows coexist and new ones drop in without touching the server. It ships seeded with a short-read variant-calling pipeline and an Oxford Nanopore long-read assembly pipeline.

You can run it two ways. Locally, with uvx flowproof-mcp, so your data never leaves your machine and your own compute does the work, which is what most people running real genomic data will want. Or against a hosted endpoint for a zero-install try. The hosted path uses per-user API keys rather than a shared secret, and the pipeline command builder is hardened against argument injection, because a service that executes pipelines on behalf of an AI has to assume the input is untrusted.

Why this matters beyond one pipeline #

There has been a call, in the bioinformatics literature this year, for MCP-enabled services so that LLM-driven discovery has a standard, machine-actionable layer to work against. FlowProof is one concrete answer to the execution-and-trust side of that vision: not another database wrapper, and not another planning agent, but the layer that actually runs the work and proves it.

It is early. It is open source. If you run Nextflow or Snakemake day to day, I would genuinely value your take on the provenance model and on which pipeline is worth wiring next.

• PyPI: pip install flowproof-mcp (or uvx flowproof-mcp)
• Source: https://github.com/ajibadedapo/flowproof-mcp