Best AI Agent Skills for Bioinformatics in 2026

Most agent-skill directories list bioinformatics tools by star count, which tells you nothing useful, because the three repositories worth installing do completely different jobs and only one of them will match what you are stuck on this week.

One of them fetches. One of them computes. One of them knows which R package to reach for. Installing all three because they all say "bioinformatics" gives you an agent with a crowded context window and no clearer idea of what it is doing.

Here is the split, with install commands checked against each repository's own README on 22 August 2026 rather than copied from a directory.

The short list

Star counts are from the GitHub API on the publication date. All three were pushed to within the last two months, which in this category is worth checking, because abandoned bioinformatics tooling is the norm rather than the exception.

BioMCP — the agent stops guessing at gene facts

BioMCP, from Genomoncology, is an MCP server that connects an agent to roughly 30 public biomedical databases. Literature through PubMed, PubTator3, Europe PMC and Semantic Scholar. Variants through MyVariant.info, ClinVar, gnomAD, CIViC, OncoKB and cBioPortal. Genes through MyGene.info, UniProt, Reactome, GTEx and the Human Protein Atlas. Trials through ClinicalTrials.gov and NCI CTS. Drugs, diseases, pathways and adverse events through MyChem, ChEMBL, OpenTargets, Monarch, KEGG and OpenFDA.

That list is the point. Ask a model unaided about a variant's population frequency and you get a number with the shape of an answer. Ask it through BioMCP and it queries gnomAD and returns the actual figure with its source.

The install expects the biomcp binary present first, then goes through the plugin marketplace:

claude plugin marketplace add genomoncology/biomcp
claude plugin install biomcp@biomcp

Most commands run without credentials. Optional keys raise rate limits or open specific sources: NCBI_API_KEY, S2_API_KEY, ONCOKB_TOKEN, plus DisGeNET, OpenFDA, NCI and AlphaGenome. Start without them.

The tool grammar covers search, get, discover, enrich and batch, with cross-entity helpers like variant trials and drug adverse-events. Those composites are where it earns its keep: going from a variant to the trials recruiting on it is three tabs and a lot of copy-pasting by hand.

MIT licensed, pushed the day before this was written.

ClawBio — the one that runs your data

ClawBio describes itself as a bioinformatics-native skill library, local-first and reproducible, and unusually for that kind of claim the repository backs it up.

It ships 96 skills, 90 with working demo data, across pharmacogenomics and CYP450 profiling, variant annotation and clinical classification, ancestry analysis with PCA, polygenic risk scoring, UK Biobank field search, GWAS lookups across nine databases, and a Galaxy bridge that reaches 8,000-plus tools. The sequencing side is wrappers around nf-core: RNA-seq for bulk transcriptomics, scRNA for single cell, Sarek for germline and somatic variant calling.

claude plugin marketplace add ClawBio/ClawBio
claude plugin install clawbio

You can also clone the repository and open it as your working directory in Claude Code, which is the better route if you want to read a skill before running it.

Python 3.11 or later, via pip or conda. The core stack is biopython, pandas, numpy, scikit-learn and matplotlib. Some skills need external binaries you install separately, Kraken2 and RGI for metagenomics, the nf-core pipelines for sequencing. That is not a gap, it is what running real pipelines involves.

The feature I would keep even if the rest disappointed is the reproducibility bundle. Skills that generate reports export commands.sh to replay what ran, environment.yml snapshotting dependencies, and checksums.sha256 over the outputs. Six months later, when a reviewer asks how a figure was produced, you have an answer that does not depend on remembering a chat session.

On licensing, be precise: the README states MIT, and GitHub's license detector returns no assertion, meaning the file is not in a form it auto-classifies. For personal use that is a footnote. For anything institutional, read the file before your compliance office does.

BioMate Bioconductor KB — for people who live in R

Bioconductor's problem is not documentation, it is that there is too much of it and the packages overlap. Six ways to do differential expression, all defensible, all with a vignette that assumes you already picked.

BioMate encodes 200 Bioconductor packages as skills: the top 100 by downloads plus 100 rising packages from 2021 onwards, filtered to analysis tools. Across those sit 390 workflows, with 89 packages carrying more than one recipe, organised into 12 domains. Transcriptomics dominates at 97 packages, then genomics at 33, general utilities at 19, proteomics at 16, epigenomics at 10.

Each skill is vignette-grounded: when to reach for the package, which parameters matter, how to read the output, what goes wrong. That is a different thing from having the model recall a function signature, and it shows in the answers.

Install is manual, and the README gives it directly:

git clone https://github.com/bioMate-AI/biomate-bioconductor-kb.git
find skills -name "SKILL.md" | while read f; do
  pkg=$(dirname "$f" | xargs basename)
  cp "$f" ~/.claude/skills/bioconductor-${pkg}.md
done

Skill content is CC-BY-4.0, the extraction scripts are Apache-2.0, and the underlying packages keep their own Artistic-2.0 or GPL terms. If you republish any of it, attribution is required.

Two hundred skills is a lot of context. Copy the domains you work in rather than the whole set.

Three install routes, and why that keeps catching people out

Every one of these repositories installs differently, and none of them uses the command that skill directories most often print.

There is no claude skill add. It does not exist. It appears across a surprising number of listing sites because it reads like it should, and running it gets you an error rather than a skill.

What does exist: claude plugin marketplace add owner/repo followed by claude plugin install name@marketplace, which is what BioMCP and ClawBio use. The npx skills@latest add owner/repo CLI, which other maintainers prefer. Copying a SKILL.md into ~/.claude/skills/ by hand, which is what BioMate documents. And claude mcp add for MCP servers, a different mechanism again.

If you want the underlying model, our explainer on Claude skills vs MCP vs GPTs covers why these are separate systems. The practical rule is shorter: read the repository README, ignore the directory.

The governance part nobody puts in the demo

ClawBio's local-first design means your VCF stays on your disk unless a skill calls out. Worth having, and it does not make the setup compliant on its own.

The prompts still travel. If you paste a patient identifier, a sample sheet with names in it, or an excerpt of clinical annotation into the chat, that content goes to the model provider under whatever terms your plan carries. Consumer tiers are excluded from the enterprise agreements that regulated work needs, and a skill that runs locally does not change what the surrounding conversation transmits.

For identifiable human genomic data, the questions are the ordinary ones: which agreement governs the model, what the retention terms are, and whether your IRB or data access committee has been asked about agent-assisted analysis at all. Most have not, which is a conversation to start rather than a reason to stop.

Public reference data, published summary statistics and your own de-identified counts matrices carry none of this weight. That covers most day-to-day work.

What these do not fix

An agent will run a pipeline with defaults that are wrong for your library prep, produce a figure that looks right, and say nothing. That is the recurring failure, and none of these three prevent it. ClawBio's reproducibility bundle at least makes the error findable afterwards, which is more than most tooling offers.

These skills remove typing and lookup from people who already know which parameters matter. They do not supply the judgement. Someone who cannot distinguish a batch effect from biology now gets to a plausible plot faster, and is no closer to a defensible result.

Used inside that limit, the time saved is real. BioMCP alone replaces a browser session of cross-referencing per variant, and that adds up over a curation shift.

For adjacent workflows, our best AI agent skills for data scientists piece covers the general analysis stack, and the full tool list for data scientists is the non-agent equivalent.

FAQ

Is there a claude skill add command for installing these? No. There is no claude skill add command, and directories that print one are guessing. Skills reach Claude Code three ways: the plugin marketplace flow, which is claude plugin marketplace add owner/repo followed by claude plugin install name@marketplace; the npx skills CLI, which some maintainers prefer; or copying a SKILL.md file into ~/.claude/skills/ yourself. MCP servers are separate again and use claude mcp add. Check the repository README rather than a directory listing.

Can I use these skills on patient or identifiable genomic data? ClawBio is designed for it in the sense that it runs locally by default and your files stay on your machine unless a skill explicitly calls an external API. That is a property of the skill, not of the model, and the prompts and any excerpts you paste still travel to the model provider. Identifiable genomic data under HIPAA, GDPR or a data use agreement needs the same enterprise contract and retention terms as any other regulated workload, plus your IRB or data access committee's sign-off on agent use.

What is the difference between BioMCP and ClawBio? BioMCP is an MCP server that fetches: it queries roughly 30 public biomedical databases including PubMed, ClinVar, gnomAD, cBioPortal, ClinicalTrials.gov and OpenFDA, and returns structured results. ClawBio is a skill library that computes: it gives the agent recipes for running analyses on your own files, including nf-core RNA-seq, single-cell and Sarek variant-calling wrappers. One answers what is known, the other processes what you generated.

Do these replace a bioinformatician? No, and the failure mode is specific. An agent will run a pipeline with defaults that are wrong for your library prep, produce output that looks correct, and never flag the mismatch. These skills remove typing and lookup time from people who already know which parameters matter. Someone who cannot tell a batch effect from a biological signal will get a plausible plot faster and be no closer to a defensible result.

Does BioMCP need API keys? Most of it works without any. Optional keys raise rate limits or unlock specific sources: NCBI_API_KEY for PubMed and PubTator throughput, S2_API_KEY for Semantic Scholar, ONCOKB_TOKEN for OncoKB therapy evidence, plus DisGeNET, OpenFDA, NCI and AlphaGenome keys for those sources. Start without keys, add them when you hit a limit.