A verifiable evidence digest for clinicians.

▶ Demo: https://youtu.be/RmT7JjTTqeM · Live: https://verastar.netlify.app

Vera (truth) + star (your north stars): verified evidence, aimed at the concepts you steer by.

You tell Verastar your north-star concepts and your active research projects. Every morning it hands you a digest of the last 1–3 days of literature that actually moves those north stars forward. Open the PubMed page, or the full-text PDF when the paper is open access. If a paper earns its salt, save it into your Library. The Connections view then surfaces the links between what you've read and the projects you're working on. And everything you save is written as plain markdown into a folder you own — your reading and your memory, on your own machine, never locked inside an app.

It is the app version of a system I already run by hand: a personal knowledge base, a context portfolio, and a daily research briefing. Verastar makes it something any clinician can pick up — and I built it end-to-end with Claude Code, as a physician who never wrote a line of code before late 2025.

Built for Built with Claude: Life Sciences (Anthropic × Gladstone Institutes), July 7–13, 2026 — Build track.

Most "AI research digest" tools ask you to trust the model. Verastar does the opposite. Two non-negotiable rules govern everything:

1. Quote-verification is the sacred core — built and protected first. Every quote, value, or citation the digest surfaces carries a verified badge. The badge is proven by the app, by independently matching the quote against the real source text — never asserted by the model. If the app can't find the quote in the source, there is no badge.

2. Never fabricate a value or a source. Flag instead. If a claim can't be tied to verifiable source text, Verastar does not smooth over it or invent a plausible citation. It flags the gap and shows you exactly what it could and couldn't confirm.

This is the whole point. Provenance is the product, not a footer.

1. Set your north stars. Define the concepts you're steering by, plus any active projects. This is your steering profile (the app version of a context portfolio).
2. Tune your digest rubric. What counts as "worth your salt" — study design, sample size, relevance to a north star, novelty — with journal preferences stored as compact structured tiers. You set the bar.
3. Daily digest. Each morning: papers from the last 1–3 days that advance your north stars, ranked by your rubric. Each item links to the DOI and offers the PDF when access allows.
4. Save what earns it. One click deposits a paper into your Library with its verified extractions attached — and writes it as plain markdown to a folder you own.
5. Connections. The Connections view surfaces throughlines across the week's reading, your existing library, and your active projects.

Verastar runs on your Anthropic API key. You paste it in; the app uses it to do the work. No shared model bill, no lock-in, and your key and reading stay yours.

Verastar is a static React + Vite app — no backend, nothing to deploy.

git clone https://github.com/MarissaFamularo/verastar.git
cd verastar
npm install
npm run dev

Open the printed localhost URL, open Settings, and paste your Anthropic API key. That's it — the digest, verifier, and library all run in your browser on your key. Chrome or Edge recommended: the "file your library to disk" feature uses the File System Access API, which Safari and Firefox don't support.

Requires Node 18+. Run the test suite with npm test (181 tests, including adversarial false-verify cases).

Cloning and running needs zero setup beyond npm install — with no configuration, Verastar is fully local: everything stays in your browser's IndexedDB, no account, no backend, and the app shows no sign-in UI at all.

To run your own instance with accounts and cross-device sync:

1. Create a free Supabase project.
2. Apply supabase/migrations/20260721000000_kv_events_with_rls.sql (paste into the SQL editor, or supabase db push) — that one file is the entire backend schema, row-level security included.
3. Copy .env.example to .env and set VITE_SUPABASE_URL and VITE_SUPABASE_ANON_KEY (Project Settings → API — use the publishable/anon key; it is safe in the bundle because RLS is the security boundary. Never put a service_role key in env or client code).

Signed in, your steering profile and library live in your account so they work on every device — and the disk folder remains your own plain-files copy. Your Anthropic and NCBI credentials remain browser-only and are never stored server-side.

On a configured deployment, setting up a real digest begins with passwordless email authentication so new libraries are cloud-backed from their first saved record. The populated demo remains account-free and separate. Libraries created locally before this account-first flow are prompted on their next visit to sign in and move their existing browser data into the account; the move remains explicit and never uploads API keys or the device-local disk-folder handle.

Built during Built with Claude: Life Sciences (July 7–13, 2026) and live at verastar.netlify.app — bring your own Anthropic key to run it. Design + planning docs:

• docs/BUILD_PLAN.md — scope, build order, never-cut list, demo script.
• docs/VERIFICATION_SPEC.md — the sacred core: the gate, algorithm, tiers, integrity rules.
• docs/EVAL.md — proving the verifier never lies: precision-first metric, ground-truth sources.
• docs/FACTS.md — locked API + data facts and the demo corpus.