vaultlab 0.0.1

From microscope to manuscript, in one repo. The AI lab for biological researchers.

vaultlab

"From microscope to manuscript, in one repo."

vaultlab is a research companion for biological scientists. Most AI lab tools take a research question and try to write the paper for you. vaultlab is different — it accompanies you through whatever you're actually doing today: searching literature, analyzing your CODEX run, drafting the methods section, building tomorrow's lab-meeting deck, triaging your inbox for the manuscript-deadline email you've been avoiding. With full context of your work — your knowledge base, your Google Docs, your Outlook calendar — Claude Code becomes a useful colleague instead of a generic chatbot.

Open-source. Local-first. Claude-Code-native. MIT licensed.

🚧 Alpha software. vaultlab is under active development toward v0.1.0 (target: late May 2026). Expect rough edges. See docs/KNOWN_LIMITATIONS.md.

What it does

📄	Literature search & citation verification — PubMed, Semantic Scholar, CrossRef, bioRxiv, Springer, Elsevier, paperclip MCP
🧬	Wet-lab data analysis — CODEX, MALDI, Visium, scRNA-seq, H&E, flow cytometry
📊	Publication-quality figures with corpus-backed recipes (every recipe cites ≥3 published examples)
✍️	Manuscript drafting with NotebookLM-style evidence retrieval — every [N] shows the exact passage on hover
🎤	Slide decks built from research outputs — journal-club, thesis-committee, conference-talk modes
🧠	Knowledge base (Obsidian-native) that links it all, queryable via semantic search
📥	Email + calendar context — vaultlab reads your Outlook (Windows) or Gmail to know what's pressing
📝	Google Docs integration — your lab work log + Sheets data + Drive files become first-class context

Companion mode, not autonomous mode

vaultlab is not an autonomous AI scientist. It does not generate experiment ideas in a vacuum, run robots, or submit papers without you. It assumes:

• You have ideas — vaultlab amplifies them
• You have context — vaultlab indexes it
• You make the calls — vaultlab does the rote 60% so you can focus on the insightful 40%
• You ship the paper — vaultlab drafts, verifies, formats, but the byline is yours alone

The "research companion" framing is intentional. The published-paper-via-AI bans many journals impose? Not our use case. "Here are 23 things vaultlab made my week easier" is.

Install

git clone https://github.com/bobbyni819/vaultlab && cd vaultlab
pip install -e ".[all]"
vaultlab setup            # interactive: API keys, KB path, Obsidian, Google, Outlook

Or, if you only want a piece (citations, lit search, figures):

pip install vaultlab            # core
pip install "vaultlab[research,citations]"   # specific subpackages

5-minute Hello World

vaultlab demo pbmc3k

In ~2 minutes on a laptop, this:

1. Downloads the 3k PBMC dataset (50 MB)
2. Runs QC + normalization + Leiden clustering
3. Auto-annotates clusters via LLM (with hedged voice and quoted evidence)
4. Renders 3 publication-quality figures
5. Builds a 5-slide journal-club deck with speaker notes
6. Auto-writes a KB summary note linking everything

Use cases (real ones, not benchmarks)

These are the workflows vaultlab solves end-to-end:

• "I have a CODEX run. Get me to a labeled figure." Ingest TIFF → segment with Cellpose → cluster → LLM-annotate → publication-tight spatial overlay → caption draft → KB note.
• "Draft me a Methods paragraph for the lung paper." Reads project KB → drafts → verifies every citation semantically → flags any HALLUCINATED → produces a draft you edit, not write from scratch.
• "Find papers using GPR55 in intestinal epithelium." Multi-source lit search (PubMed + bioRxiv + paperclip MCP) → smart query expansion → dedupe → re-rank → KB ingest of top 10 → citation-graph view.
• "Build me a journal-club deck on Smith et al. 2024." /paper-to-slides 10.1038/... extracts figures from PDF → composes 12-slide deck → auto-drafts speaker notes → exports .pptx.
• "What's on my calendar this week + which manuscripts are due?" Outlook reads upcoming meetings, Gmail reads journal deadlines, KB cross-references active manuscripts → integrated daily brief.

See docs/use-cases.md for more (post-v0.1).

Architecture philosophy

vaultlab is a capability layer FOR Claude Code, not a competing harness. Markdown is the user-facing interface; Python is the engine. Slash commands, role prompts, recipes, layouts, and skill definitions are all markdown files Claude Code can read at first repo open.

See docs/architecture.md for the full spec.

The four core commitments

1. Markdown is the interface; Python is the engine. Slash commands, role prompts, workflow descriptions are markdown.
2. Anti-laziness on semantic reading. Every LLM call requires quoted evidence. No surface-skim.
3. Result-oriented agentic loop. User says "draft methods" → vaultlab plans + verifies + refines internally → returns finished result.
4. KB is the smartness. Every analysis writes to KB; every analysis reads from it. The LLM gets smarter project-by-project.

What's unique vs PaperQA / scanpy / FutureHouse / scverse / Aider

	vaultlab	PaperQA2	scanpy	FutureHouse	scverse	Aider
Wet-lab data analysis	✓	✗	✓	✗	✓	✗
Literature + citation verification	✓	✓	✗	✓	✗	✗
NotebookLM-style evidence retrieval	✓	partial	✗	✗	✗	✗
Manuscript drafting	✓	✗	✗	partial	✗	✗
Slide deck output	✓	✗	✗	✗	✗	✗
Calendar / inbox context	✓	✗	✗	✗	✗	✗
Knowledge base (Obsidian)	✓	✗	✗	✗	✗	✗
Local-first	✓	✓	✓	✗	✓	✓
Companion mode (not autonomous)	✓	partial	n/a	✗	n/a	✓
Claude-Code-native skill bundle	✓	✗	✗	✗	✗	partial

No tool does all of these. vaultlab's value is the combination — wet-lab analysis (scverse-grade) + literature verification (PaperQA-grade) + manuscript + slides + life-context (calendar/inbox/docs) wired through Claude Code.

If you only need one piece, those tools are great. If you want a research companion, vaultlab is the only OSS option.

See docs/comparison.md for the full positioning analysis.

Demos

Demo	Dataset	Time
examples/pbmc3k	3k PBMCs (scRNA-seq)	2 min — Hello World
examples/visium_brain	10x mouse brain Visium	30 min — spatial transcriptomics
examples/codex_hubmap_tonsil	HuBMAP tonsil CODEX	30 min — flagship spatial imaging

Documentation

Setup:

• docs/setup-obsidian.md — Obsidian + plugin walkthrough
• docs/setup-api-keys.md — Anthropic + literature API keys
• docs/setup-google.md — Google ecosystem (Docs, Sheets, Drive, Gmail, Calendar)
• docs/setup-outlook-windows.md — Outlook COM (Windows-only)

Reference:

• docs/architecture.md — full architectural spec
• docs/use-cases.md — concrete examples of what vaultlab solves
• docs/comparison.md — vs other tools

Privacy & limits:

• docs/data-privacy.md — what data leaves your machine
• docs/compliance.md — explicit non-HIPAA disclosure
• docs/long-term-reproducibility.md — model-version philosophy
• docs/KNOWN_LIMITATIONS.md — honest failures

For contributors:

• CONTRIBUTING.md — how to contribute
• AGENTS.md — invariants and conventions
• CLAUDE.md — entrypoint for Claude Code sessions
• INSPIRATIONS.md — what we drew from where (auditable lineage)

Citation

See CITATION.cff. Once v0.1.0 ships, the preferred citation is:

@software{ni_vaultlab_2026,
  author = {Ni, Bobby Y.X.},
  title  = {vaultlab: A research companion for biological scientists},
  year   = 2026,
  url    = {https://github.com/bobbyni819/vaultlab},
  version= {0.1.0}
}

Privacy & compliance

vaultlab uses Anthropic's Claude API. Prompt content is sent to Anthropic. vaultlab is NOT HIPAA-compliant. Do NOT use with PHI/PII/IRB-restricted data. See docs/data-privacy.md.

When you opt into Google or Outlook integration, vaultlab also reads:

• Google Docs / Sheets / Drive content you authorize
• Gmail messages matching your search criteria
• Outlook calendar events + email subjects/bodies

This data may end up in prompts sent to Anthropic. Do not enable Google/Outlook integration if your account contains PHI or institution-restricted data. Each integration has its own scopes you can audit; see docs/data-privacy.md.

By using vaultlab, you take full responsibility for compliance with your institutional, IRB, IACUC, and regulatory obligations.

Author

Bobby Y.X. Ni — Hickey Lab, Duke University Biomedical Engineering.

License

MIT — anyone can use, modify, distribute, including commercial.