Curated, evidence-grounded skill and software-tool collections for scientific AI agents, generated by the AgenticScienceBuilder (ASB) pipeline. Each skill is distilled from a peer-reviewed method paper and its public code repository, anchored to verbatim evidence, EDAM-annotated, and gated for licensing/PII before release.
This is a community effort, initiated at the Dagstuhl Seminar Computational Metabolomics (26181, 26–30 April 2026).
This release —
metabolomics-v0.1.0(preliminary):collections/metabolomics/v2— 5,865 skills across 909 tools distilled from 568 papers, for computational metabolomics — predominantly LC-MS/MS, but also LC-MS, GC-MS, mass-spectrometry imaging, ion mobility and lipidomics, with some NMR and multi-omics / statistics / pathway tools.
| Artifact | Status |
|---|---|
| ASB-Skills — evidence-grounded procedural skills | ✅ released |
| ASB-Tools — deduplicated software-tool records (EDAM + DOIs) | ✅ released |
| ASB-Benchmark — per-paper tasks + claim-retrieval test sets | ⏳ to be released soon |
| ASB-Capsules — raw per-paper ASB pipeline outputs (full traceability) | ⏳ to be released soon |
Only ASB-Skills and ASB-Tools are published now; the benchmark and capsule layers follow (see PROVENANCE.md).
asb-skill-collections is a multi-domain marketplace — it hosts ASB-generated skill
collections for any scientific domain, organized by provenance (ASB-generated), not by
field, so the repo name and marketplace stay domain-agnostic as new domains are added:
collections/<domain>/<version>/ # full collection per domain (e.g. metabolomics/v2)
packs/<domain>/<technique>/ # lighter per-technique subsets (e.g. metabolomics/lc-ms)
Each domain ships a full plugin (<domain>) plus per-technique packs (<domain>-<technique>).
Metabolomics is the first released collection; proteomics, transcriptomics, epigenomics and
further domains follow under the same layout — no rename, just new entries under
collections/ + marketplace.json.
Tip
Fastest path — two lines in Claude Code: the full collection, or a lighter per-technique pack.
/plugin marketplace add HolobiomicsLab/asb-skill-collections
/plugin install metabolomics@asb-skill-collections # full collection (5,865 skills)/plugin install metabolomics-lc-ms@asb-skill-collections # also: gc-ms, nmr, ms-imaging,
# ion-mobility, ce-ms,
# direct-infusion, ms-genericNote
Packs overlap (a multi-technique skill appears in several) — install one full plugin or a few packs, not both. See packs/metabolomics/.
No CLI: upload the search indexes + the few skills you need as the assistant's knowledge (Claude Projects, ChatGPT Custom GPT/Project, Mistral Agent/Library) and paste a routing instruction — step-by-step in USAGE.md.
The collection is plain Markdown + JSON — point your
agent at collections/metabolomics/v2/ and read the indexes. See
AGENTS.md.
Other agent runtimes have no /plugin install. Use the bundled asbb CLI from a
local clone to materialize a pack into the runtime's own location.
git clone https://github.com/HolobiomicsLab/asb-skill-collections.git
cd asb-skill-collections
python3 -m scripts.asbb_cli install --list-runtimes # see all targetsSkill-native runtimes (read SKILL.md directly):
# Codex + Copilot CLI + Gemini CLI all share ~/.agents/skills — one install:
python3 -m scripts.asbb_cli install metabolomics-lc-ms --runtime agents
# Or a specific home: --runtime codex | copilot | gemini
# Vendor into a project for Claude Code: --runtime claude (add --user for ~/.claude)Rules/instruction IDEs (a SKILL.md is rendered into their format — run from
the target project):
python3 -m scripts.asbb_cli install metabolomics-lc-ms --runtime cursor # .cursor/rules/*.mdc
python3 -m scripts.asbb_cli install metabolomics-lc-ms --runtime cline # .clinerules/*.md
python3 -m scripts.asbb_cli install metabolomics-lc-ms --runtime vscode-copilot # .github/instructions/*.instructions.mdAnything else (pi, Antigravity, or a runtime without a preset):
python3 -m scripts.asbb_cli install metabolomics-lc-ms --dest ~/some/skills/dirSkill-native installs symlink by default (a git pull in the clone updates
them); add --copy for a self-contained copy. --dry-run previews, --force
overwrites unmanaged files, and asbb uninstall <pack> --runtime <id> cleanly
removes exactly what was installed (tracked in ~/.asbb/installed.json).
For Claude Code, the plugin marketplace above remains the recommended path.
Search → apply → ground. Find a skill via skills_index.json (by EDAM IRI,
tool name, or keyword) or tools_index.json; read its SKILL.md and follow the
procedure; then optionally ground it against the source paper/repo to verify a
parameter or claim — see the 🔎 Grounding (Perspicacité) section below.
Requirements (libraries, per-skill tool deps) are in USAGE §0.
Skills carry distilled procedure; to verify an exact parameter, threshold, or claim, ground a skill against the paper it was distilled from. Grounding is optional and additive — every skill works without it — and ships inside every plugin and pack.
It's powered by Perspicacité — Holobiomics Lab's local-first scientific literature-RAG engine. Two backends, KB-primary with a serverless fallback:
kb(Perspicacité) — RAG over the source paper's full text + supplementary information, persistent and citable. The per-paper KB (asb-paper-<doi>) is auto-created and ingested on first use via the MCP toolsensure_kb/ground_paper.local(serverless) — no server:git clonethe skill's source repo + best-effort open-access paper, then read the files directly.
Use it
- In Claude Code: run
/groundon the skill in play — it installs + queries the source KB, or falls back to a local clone when no server is running. - Anywhere: call the bundled
bin/perspicacite_kb_bind.py(prepare/query/local).
Tip
Recommended before you act on a numeric parameter, threshold, or quantitative claim — it's the difference between "the skill says ~5 ppm" and "the paper specifies 5 ppm."
The kb backend needs a reachable Perspicacité (PERSPICACITE_BASE, default
http://127.0.0.1:8000); the local backend needs only git + network. Full guide:
USAGE.md §4.
The LC-MS view of the collection (the metabolomics-lc-ms pack): papers → ASB → EDAM-typed skills, routed by Perspicacité. The full release spans 5,865 skills across all techniques.
| File | Contents |
|---|---|
skills/<slug>/SKILL.md |
one evidence-grounded skill each (frontmatter: EDAM IRIs, derived_from DOIs, evidence_spans, tools, attribution) |
tools/<slug>.yaml |
deduplicated software-tool records with EDAM + source DOIs |
skills_index.json / tools_index.json |
machine search indexes |
kb_bundle.json |
skill → source-paper KB slugs + repo_urls (grounding map) |
bin/perspicacite_kb_bind.py · commands/ground.md · GROUNDING.md |
packaged grounding — the binder, the /ground command, and a how-to (shipped in every plugin & pack) |
collection.yaml · corpus.yaml |
collection record · per-paper access basis (repo-oa) |
CITATION.cff · PROVENANCE.md · gate_report.json |
citation · how it was generated · release-gate verdict |
The default entry point is skills/_router/SKILL.md.
For a description of the collection's content (technique & EDAM-topic breakdown) and how it was selected (sources, inclusion/exclusion criteria, grounding, gating), see ABOUT.md.
The exact ASB build command and the mixed-model routing (Opus 4.8 for
outline/card-revision, Haiku 4.5 for the rest, OpenAI embeddings) are documented
in PROVENANCE.md, recorded per build
in build_manifest.json. The raw per-paper ASB capsules and the benchmark
layer (full end-to-end traceability) will be released later.
If you use this collection, cite both the collection and the original
paper behind each skill you use (attribution.original_doi).
- Collection authors (Zenodo, see
CITATION.cff): AgenticScienceBuilder Community, Louis-Félix Nothias, HolobiomicsLab.cnrs.fr, MetaboLinkAI.net. - Per-skill roles (in each
SKILL.mdattribution:block):generator(the ASB pipeline) ·curators(who modify/validate — none yet) ·promoter(suggests use — Louis-Félix Nothias) ·sponsor(paid the API cost — CNRS & Université Côte d'Azur) ·original_doi(source paper). - Zenodo DOI: 10.5281/zenodo.20794027.
This collaborative project was initiated at the Dagstuhl Seminar Computational Metabolomics (26181, 26–30 April 2026), and we thank the seminar's participants and organizers. API generation costs for this collection were sponsored by CNRS and Université Côte d'Azur. Built with the AgenticScienceBuilder pipeline and grounded with Perspicacité (Holobiomics Lab).
Skills are distilled from peer-reviewed method papers anchored to the
computational metabolomics review series (Misra → Enveda). See
governance/SOURCES.md for the source inventory and the
scientific inclusion criteria, and governance/CONTENT_POLICY.md
for the legal/open-access policy.
Suggest or annotate a paper: see Contributing → Propose or annotate a paper.
Every skill carries a license_tier field (in skills_index.json and in each
SKILL.md frontmatter metadata.license_tier) that answers what may I do with
the underlying tool?
| Tier | Meaning |
|---|---|
open |
Commercial use OK (MIT, Apache-2.0, GPL, CC-BY, …) |
noncommercial |
Academic / noncommercial only — confirm permitted use before applying the skill |
restricted |
No clear license detected — verify before commercial use or redistribution |
Discovery defaults to open skills; the asb-metabolomics meta-skill enforces
the noncommercial acknowledgment gate. Non-open skills carry a one-line banner
in their body. Full policy: governance/LICENSE_TIERS.md.
# list only open-tier skills
jq '[.[] | select(.license_tier=="open")]' collections/metabolomics/v2/skills_index.jsonDual-licensed, by layer (see LICENSING.md):
- Code (scripts, tooling) — Apache-2.0.
- Collection content (skill descriptions, tool records, structured metadata) —
CC-BY-4.0 (as stamped in every
SKILL.md,collection.yaml,CITATION.cff). - Verbatim quotations from source papers (
evidence_spans) — minimal, attributed, under fair-use / quotation right; a rights holder may request removal.
Maintained by Holobiomics Lab — see
MAINTAINERS.md. Curator workflow: CONTRIBUTING.md;
conflict-of-interest policy: COI_POLICY.md. All governance & policy
docs now live in governance/.
collections/ also contains epigenomics/v1, transcriptomics/v1, and the
earlier metabolomics/v1. These are staged/internal and not part of this
release — only metabolomics/v2 is published via the plugin.
- Zenodo DOI — 10.5281/zenodo.20794027.
w3id.org/holobiomicslab/…IRIs — reserved identifiers that do not resolve yet (the redirect is not live); treat as stable names, not links.- HuggingFace mirror & leaderboard — planned, not yet live.
- Benchmark / capsules — to be released later.