viCRO — The open source virtual CRO

Sourcing samples is broken

You can't see where samples come from, whether they'll work for your assay, or how much they should cost. The information exists — it's just buried.

20-60x

markup from hospital vial to buyer. No published rate cards.

80%

of researchers limited their work because they couldn't get specimens.

91%

of biobanks have no public access policy. Can't find what won't publish.

<10%

utilization at over half of biobanks. Samples exist — they're just stuck.

Type vcro

This is what opens. Your wiki, your workflows, one prompt away.

vcro

╦ ╦╦╔═╗╦═╗╔═╗ ┌────────╚╗╔╝║║ ╠╦╝║ ║────────┐ │ virtual ╚╝ ╩╚═╝╩╚═╚═╝ │ │ sourcing agent v2.3.0 │ └───────────────────────────────┘ 338 wiki entities · /source /compile /bounty /onboard /lint ❯ I need stool samples from IBD patients to run shotgun metagenomics, treatment-naive, at least 100 subjects. ⟡ searching wiki index ⟡ searching PubMed: IBD stool shotgun metagenomics ⟡ searching ClinicalTrials.gov: IBD stool microbiome ⟡ scoring 5 candidates on Scale · Cost · Quality Can you get it? Yes, partially — SPARC-IBD and IBD BioResource both hold stool at scale, but treatment-naive N is unconfirmed for every candidate. [inferred] Best path: SPARC-IBD (CCF) → Weill Cornell Microbiome Core — ~7,000 subjects, $250/sample, ~$25K for n=100. Treatment-naive subset inquiry is the next step. What blocks it: No candidate has confirmed its treatment-naive specimen count. [open_question] ┌─────────────────────────────────────────────────────┐ │ Path │ N │ Assay (n=100) │ Timeline │ ├─────────────────────────────────────────────────────┤ │ SPARC → Cornell ~7,000 $25,000 10–18 wk │ │ SPARC → MDA ~7,000 $32,500 10–18 wk │ │ BioResource 25,000 ~$25K+ship 14–22 wk │ └─────────────────────────────────────────────────────┘ 47s · a3f8b2c1 ❯ █

You ask. The graph grows.

Not a search engine. A knowledge graph that compounds with every question.

◉ Reads papers, not abstracts

8,000 words per paper. 21 intelligence dimensions extracted — sample counts, protocols, consent, PI contacts, platform specs. With verbatim quotes.

◎ Cross-links everything

Every cohort linked to its institution, investigators, platforms, protocols. Ask about a PI — see every cohort they touch. Ask about a platform — see who validated it.

◆ Scores on three axes

Scale, Cost, Quality — independently. No composite rank. You decide the weighting. Every score traces to a source quote.

▸ Self-improving

Lint scans the wiki for stale facts, missing links, duplicate entities. Findings feed back into compile. The graph heals itself.

Five workflows

Natural language or structured commands. Every answer backed by evidence.

vcro source "AD plasma metabolomics cohorts, longitudinal, n>200": Search the wiki. Score matches. If the wiki is thin, ingest new papers automatically. Get a cited recommendation.
vcro compile PMC10103184 PMC6922070: Your wiki is only as good as what it has read. Compile turns papers into structured entities — cohorts, institutions, investigators, platforms — all cross-linked.
vcro bounty "50 AD plasma samples, commercial use, under 80k EUR": Need physical samples? Bounty maps a procurement chain (specimen source → provider → assay → delivery), prices each link from verified sources, and flags every component that still needs a quote.
vcro lint: The wiki audits itself. What facts went stale? Which entities are duplicates? Where are links missing? Findings feed back into compile — the graph heals.
vcro onboard "Sahlgrenska Biobank": Supply side. Generates a catalog listing from an institution's published papers. Surfaces gaps for the biobank to fill. Same evidence standard, both sides.

Skills, not features

21 documented instruction sets under .claude/skills/. Each one is a markdown file. Read it, fork it, swap it out.

↓ Download skills package

compileextract

One paper → structured intelligence fragments. 5–8 dimensions.

compileresolve

Fragments vs. wiki. New entity, merge, or ambiguous.

compilemerge

Plan in, wiki articles out. Idempotent. Hook-gated.

queryunderstand

Natural language → structured request.json.

querysearch

PubMed, EuropePMC, ctgov. Coverage scoring per round.

querydiscover

Wiki-first candidate search. Reports gaps.

queryscore

Scale, Cost, Quality. Per-axis confidence. No composite.

querydeliver

Scored candidates → recommendation.md + listings.jsonl.

querybounty/format

Procurement chain → wiki bundle entity. Per-link evidence.

lintgaps

Real gaps vs. structural. Worth recompiling or not.

lintconsistency

Duplicate cohorts, distinct waves, or genuinely separate.

lintstaleness

Needs re-verification vs. static facts that aged.

lintconnections

Latent links to apply. Papers worth ingesting.

catalogcatalog

Institution listing draft from wiki entities.

catalogcompliance

Consent, commercial-use, export, IRB scope.

catalogprice

Cost-recovery pricing from verified analogues.

infraagent-endpoint

Query API for external agents. Dynamic per run.

infrajobs

Batch runs, compile wedges, lint sweeps.

infracontributing

How to add a skill without breaking the lockfile.

infrawebapp

Next.js + Vercel AI SDK + Claude bootstrap.

Open source. Free.

viCRO is MIT-licensed. You need a Claude subscription for the LLM calls — that's it. The wiki, the graph, the skills all live on your machine.

◇

kamilseghrouchni/vcro-sourcing

21 skills · 4 agents · 9 rules · MIT

◎

Claude Code

Runtime. Your subscription powers the agents.

⍑

vis-network

Interactive graph visualization.

⚙

Python stdlib

Zero dependencies for core scripts.