ClawBio: a local-first skill library that puts genomics tools inside your AI editor

What ClawBio solves for people who already read VCFs

Bioinformatics tooling is fragmented across command line programs, web portals and Python packages, and the glue between them is usually a notebook that nobody else can rerun. ClawBio's answer is to express each analysis as a skill that an AI agent can invoke, with a folder layout rather than a bespoke plugin API. The README describes the project as "the first bioinformatics-native AI agent skill library" and lists pharmacogenomic dosage cards, variant lookup across nine genomic databases, polygenic risk scoring and UK Biobank field discovery as the things it does today. The audience is narrow on purpose: people who know what an rsID is, who have a genome file or a cohort, and who want the analysis to stay on their own machine. The local-first framing is not decoration. The demo genome is the Corpasome, a real individual's 23andMe chip data plus subsets of a 30x Illumina whole-genome sequence, released under CC0. If you are looking for a hosted service that accepts uploads, this is the wrong shape of project.

Skills as folders, and the MCP server on its way out

The architectural shift in 0.7.0 is the interesting part. Earlier releases offered an MCP server, started with uvx --from 'clawbio[mcp]' clawbio mcp, which exposed the skills to any MCP-capable editor. The 0.7.0 release notes call that server deprecated and say it will be removed in 0.8.0, though existing configurations keep working until then and print a notice on start. What replaces it is simpler: every skill is a plain Agent Skills folder, and you copy or symlink the folders you need from skills/ into ~/.agents/skills/, which the README says Claude Code, Cursor, VS Code, Codex and Zed all read, or into a project-local .agents/skills/. No server process sits between the editor and the analysis. That is a real reduction in moving parts, and it also means the integration surface is a directory convention rather than a protocol version. The migration path is documented at docs.clawbio.ai/reference/mcp. The trade-off is that anything you built against the MCP endpoint has a deadline attached to it.

Installation paths: pip, conda, source checkout, editor plugin

The README gives four routes and they are not equivalent. The fastest is two commands: pip install clawbio on Python 3.11 or newer, then clawbio run pharmgx --demo. A conda route exists as conda install -c bioconda clawbio. For library use, the documented import is from clawbio import run_skill, list_skills, with run_skill("pharmgx", demo=True) returning a result object. If you are developing ClawBio itself or you want every skill with full demo data, the README points you at a source checkout with uv: git clone the repository, cd ClawBio, uv sync to install from pyproject.toml and uv.lock, then uv run python clawbio.py run pharmgx --demo. Claude Code users get a plugin route through /plugin marketplace add ClawBio/ClawBio. Note the distinction the README draws between the installed package and the source checkout: the checkout is what gives you all skills plus the complete demo data, so a benchmark run against a pip install may not match what the repository's own test commands assume.

Validation infrastructure is the part worth reading closely

The 0.5.0 release added a benchmark suite, and it is the most concrete thing in the material. There is an Alzheimer's disease ground truth set of 34 genes across three evidence tiers, 20 negative controls and 10 lead variants. There is a mock API server with deterministic endpoints for Ensembl, GWAS Catalog and ClinPGx, which lets CI run offline without hitting rate limits. There is a scorer that reports gene recovery rate, FDR, precision, recall, F1 and a tier-weighted composite score. Fine-mapping is swappable between ABF and SuSiE on the same data, with a method registry pattern the README says is there to add FINEMAP and PolyFun later. A nightly sweep runs every skill in demo mode and scores the outputs. The documented commands are python tests/benchmark/finemapping_benchmark.py --output /tmp/fm_bench, python tests/benchmark/benchmark_scorer.py --genes "APP,BIN1,CLU,TREM2,GAPDH" and python tests/benchmark/mock_api_server.py. The README claims 74 benchmark tests at the 0.5.0 baseline and a public leaderboard, but the text is truncated at that point, so treat the leaderboard as unverified. The AD-only ground truth is a genuine limitation: a skill outside that domain is not scored by this suite at all.

Reproducibility bundles come with an asterisk

Many analyses write a reproducibility/ directory containing replay commands, environment metadata and output checksums. The README is unusually candid that the exact files vary by skill and that some replays still require the original external inputs to be present. That sentence matters more than the feature description. A checksum proves an output has not changed; it does not prove the output can be regenerated on a different machine if the upstream query result or the input file is gone. The mock API server partly addresses this for the three services it covers, but any skill that queries a database outside Ensembl, GWAS Catalog and ClinPGx has no offline replay path described in the material. If reproducibility is your reason for adopting ClawBio, start by reading docs/reproducibility.md and confirming that the specific skill you care about produces a bundle you can actually replay.

Where ClawBio is the wrong tool

Three cases. First, anything clinical. The README states that the reference genome data is provided for research and educational purposes only, and the warfarin example in the README is a demonstration of a genotype lookup, not a validated prescribing system. Second, anything that needs a network API. The MCP server is deprecated, so if your architecture assumes a long-running service that other tools call, you are building on a component with a stated removal date in 0.8.0. Third, anything outside the skills that ship. The project advertises 97 skills and 8,182 Galaxy tools, but the validation suite scores against an Alzheimer's disease ground truth, and the README does not describe per-skill accuracy for the rest. A polygenic risk score built from thousands of small-effect variants is sensitive to the reference panel and the summary statistics used; nothing in the supplied material documents how ClawBio handles that, so you would be trusting the skill's implementation rather than a published benchmark.

How it differs from Galaxy and from writing your own scripts

The obvious comparison is Galaxy, and ClawBio's own numbers invite it: 8,182 Galaxy tools are referenced alongside the 97 skills. The difference in approach is where the orchestration lives. Galaxy is a web platform with a workflow editor, a tool wrapper format and a server you operate or rent; its unit of reuse is a tool definition and its unit of execution is a job on that server. ClawBio's unit of reuse is a folder of instructions and code that an agent reads, and its unit of execution is your local Python environment driven by a natural-language request. Galaxy wins on provenance, sharing and multi-user infrastructure. ClawBio wins on setup time and on keeping genotype data off someone else's machine. The other alternative is a hand-written Snakemake or Nextflow pipeline: more control, no agent in the loop, and no demo data to get you started. ClawBio's demo genome and mock API server are the parts a hand-rolled pipeline would take longest to replicate.

Maintenance cost and the licence question

The release cadence visible in the material is roughly one tagged release every six to eight weeks, with 0.5.2 in June 2026, 0.6.1 in July, 0.7.0 in September, and the last push to main on 2026-09-10. The 0.7.0 deprecation of the MCP server with removal in 0.8.0 is the concrete upgrade cost you should price in: if you adopted the MCP route in 0.6.x, you have one release cycle to move to the skills folder layout. The licence is the loose end. The repository metadata reports NOASSERTION, while the README badge and the badge markup both say MIT. Those cannot both be right, and the discrepancy is worth resolving before you ship anything derived from the code, since MIT and an unasserted licence carry different obligations. This is a factual observation about the repository, not legal advice; read LICENSE yourself and get counsel if the answer affects your distribution.

Editorial conclusion

Adopt ClawBio if you are a computational biologist or bioinformatician who already works in an Agent Skills aware editor and wants genomics skills to run locally against your own data, and if you can accept a young project whose licence file does not match its MIT badge. Do not adopt it if you need a stable remote API, since the MCP server is deprecated and slated for removal in 0.8.0, or if you need a supported clinical decision tool rather than a research and educational one. Before committing, run pip install clawbio and clawbio run pharmgx --demo on your own machine, read LICENSE to resolve the NOASSERTION status, and check whether the reproducibility bundle for your chosen skill can be replayed without the original external inputs.

Official sources

1. ClawBio/ClawBio on GitHub
2. Issues
3. Project website
4. README
5. Releases