Statistical Genomics Foundry
Design docs design record

ClawBio

A reproducibility- and benchmark-forward bioinformatics skill library on OpenClaw — its framing, strengths, and where it sits on our values.

github.com/ClawBio/ClawBio — "the first bioinformatics-native AI agent skill library. Local-first. Reproducible. Built on OpenClaw." MIT; 89 skills (29 production-ready), a Galaxy bridge over 8,000+ tools, and benchmark validation; v0.5.0 (Apr 2026). Verified 2026-06-27 against the repo README and the project site.

Its framing

ClawBio is a bioinformatics-native agent skill library built on OpenClaw, organized around three claims: local-first privacy (genomic data never leaves the laptop), reproducibility (every run can be replayed without the agent), and validation against ground truth. Each skill is a self-contained directory — a declarative SKILL.md, validated Python, demo data, and a reproducibility/ bundle. Domains span pharmacogenomics, variant interpretation, population genetics and ancestry, single-cell, and a Galaxy bridge over thousands of tools.

Strengths relative to our approach

On two axes ClawBio is ahead of most skill libraries — including bioSkills:

  • Reproducibility as a first-class artifact. Many skills emit a reproducibility/ bundle on run: commands.sh (replay), environment.yml (conda snapshot), SHA-256 checksums, and sometimes a runtime-lock.json — so a run can be "reproduced without the agent." (It is a feature of many skills, not yet universal — a handful of bundles are committed as examples.)
  • Benchmark validation with sourced ground truth. v0.5.0 adds a benchmark suite: a curated Alzheimer's-disease ground-truth set — cited to Bellenguez 2022, Wightman 2021, ClinVar, and OMIM — scoring fine-mapping output on gene-recovery / FDR / precision / recall / F1, plus a mock API server for offline CI and a nightly sweep. Empirical, not LLM self-assessment. Coverage is narrow and honestly scoped: a ~10-skill smoke leaderboard, one skill (fine-mapping) deeply ground-truthed, and by ClawBio's own rubric the blocking bench-validated tier is still unpopulated.
  • Local-first / privacy by design — data stays on the laptop; no cloud uploads; no Docker or Singularity required for core functionality.
  • Genuinely multi-runtime — a Claude Code plugin, a Python import, chat adapters, and the OpenClaw gateway, with reproducibility explicitly decoupled from any AI vendor (runs under Claude, ChatGPT, or Ollama).
  • A machine-readable catalog (skills/catalog.json) — status flags, maturity_tier, trigger keywords, and chaining partners — more structured than plain-text related-skills links.
  • Explicit data-source provenance — CPIC, GWAS Catalog, ClinVar, gnomAD, and Zenodo DOIs for reference data.

Where it sits on our values

Value ClawBio
Produces skills Yes — 89 skills (29 production-ready), each a self-contained SKILL.md + validated Python + demo data + reproducibility bundle. Exactly our packaging target.
Progressive disclosure Partial — the Agent Skills shape (declarative SKILL.md → on-demand Python/data), surfaced as a Claude Code plugin, with a catalog.json of trigger keywords. Disclosure sits within the artifact; there is no governed layer beneath it.
Traceability Strong on reproducibility — replayable commands, pinned environments, SHA-256 checksums, cited data sources and DOIs. But the SKILL.md is the source of truth, so traceability runs to external data and a replayable run, not to an authoritative upstream source with provenance lineage. The method and literature citations a skill carries are author-supplied prose within the artifact, not resolved against a governed source — so a reader cannot tell, from the skill alone, which references trace exactly and which are approximate.
CLI instructions derived from CLI docs Partial — reproducible environment pinning (uv.lock, conda, runtime locks) rather than per-tool version-compatibility / introspection derived from CLI documentation. It captures the whole environment instead of disciplining each invocation.
Portability High — Claude Code plugin, Python import, chat adapters, OpenClaw gateway; reproducibility decoupled from any AI vendor.
Human scrutiny Yes, at the file — hand-written SKILL.md + Python, PR-contributable, with a reproducibility bundle a human can replay without the agent. But the distribution site and catalog are not a navigable surface a human reads to learn the domain.
Knowledge-base backed No — the skills are the source of truth; external databases are queried per-skill and catalog.json indexes the skills, but there is no inspectable human knowledge base beneath them.

Where it diverges from the Foundry

ClawBio is the rare skill library that already does two things most do not: many skills emit a reproducibility bundle on run, and it scores a handful of skills against curated ground truth — one (fine-mapping) deeply, against a sourced AD set — in a nightly smoke benchmark, empirical rather than self-certifying. Two precise distinctions remain. First, reproducibility is not provenance: ClawBio guarantees a run can be replayed — same commands, same environment, same checksums — whereas the Foundry traces an artifact back to a governed source of truth; a reproducibly-wrong analysis still reproduces. Second, the unit validated: ClawBio's benchmark asks "did this skill's output match a known answer set?" — output-versus-truth, per-skill, offline — where the Foundry's referee asks the prior question for an arbitrary analysis, in-loop: is the method statistically valid — assumptions met, no double-dipping, the named method real and appropriate, error actually controlled. And like every hand-authored library, its SKILL.md files are the source of truth: no abstract source cast into artifacts, no foregrounded human knowledge base. ClawBio is among the strongest of these libraries; the Foundry argues for the layer beneath it.

Sources