Scholastic

A research workspace that refuses to make things up.

Search nine open scholarly sources at once, read papers with a copilot that cites the page it answered from, compare them in a matrix where every cell carries its evidence, and draft manuscripts and slide decks that cannot contain an uncited claim.

Self-hostable, MIT-licensed, and useful with no API keys at all — every AI feature is optional and says so when it is switched off, rather than breaking.

---

Why this exists

Most research tools generate confident prose and leave you to check it. Scholastic inverts that: attribution is enforced mechanically, and unattributable output is deleted before you see it.

• A drafted sentence with no citation is removed, not flagged. Same for a slide bullet. If nothing survives, you get an empty draft that explains why — never invented findings.
• A matrix cell that a paper does not report stays blank, and that is a database constraint rather than a convention. "Did not report it" and "we failed to find it" collapsing into the same blank is how a comparison table starts lying.
• Every extracted value keeps its page and its sentence. Click a number, land on the line it came from.
• Structural graph claims are labelled "within the loaded subgraph", because a centrality score over a partial citation network is not a fact about the literature.

What it does

Search across nine sources, merged

OpenAlex · Crossref · PubMed · arXiv · Semantic Scholar · Europe PMC · CORE · DOAJ · Unpaywall, queried in parallel, deduplicated by DOI with a fuzzy fallback, then ranked. Dedup favours precision: a preprint and its published version may show twice, which is a better failure than silently merging two different papers. When a source is down, results still come back and the page says which source is missing.

Read, extract, compare

A split PDF reader whose chat answers with [p. N] citations you can click. Table and statistic extraction from uploaded papers. An evidence matrix — papers down the side, fields across the top — where "Fill empty cells" runs extraction across every paper at once and each filled cell keeps its provenance. Exports to CSV for a spreadsheet or Markdown for a draft, quotes included.

Map the citations

A full-page graph explorer over a tested pure graph layer. Expand a node's citations, ask why one paper cites another, filter by year or citation count.

Write with citations that are objects, not text

A manuscript editor where citations are structured references: the bibliography derives itself, switching style rewrites nothing, and exports go to Markdown, LaTeX, BibTeX and RIS. Plus a deck outliner that turns a topic into Marp slides — one claim per bullet, each carrying its source.

A capacity commons, not a paywall

AI features run on donated capacity. Credits allocate a shared pool; they are never purchasable, never transferable, and never earned by using the app — which is why running out returns a 403 rather than a 402. Nothing here is for sale.

Sponsors can donate three ways, differing in custody rather than technology: a key the operator holds in the environment, a relay they run and never hand over (for labs whose grant terms forbid exporting a credential), or cluster hours on a schedule. The database never stores a provider key — only the name of an environment variable. There is no web form that accepts a pasted API key, and there will not be one.

Quick start

Requires Node 20+, pnpm, and PostgreSQL with the pgvector extension.

pnpm install
cp .env.example .env.local   # set at least DATABASE_URL — see SETUP.md
pnpm db:migrate
pnpm dev

Open http://localhost:3000 and search for something. No API key is needed to search — the nine scholarly sources are open. Add an LLM key later to switch on summaries, chat, extraction and drafting; without one, those features are absent and say so.

pnpm test       # 764 unit + integration tests (Vitest)
pnpm test:e2e   # 29 browser tests (Playwright) — none of them spends a credit
pnpm lint
pnpm build

Built with

Next.js 16 (App Router, Turbopack) · React 19 · TypeScript · Tailwind v4 · PostgreSQL + pgvector · Drizzle ORM · better-auth · Tiptap · pdf.js · Vitest · Playwright

Six LLM backends are supported behind one interface (Anthropic, Groq, SambaNova, Mistral, OpenRouter, Google), plus any OpenAI-compatible endpoint — including a local Ollama or vLLM server, which keeps everything on your own hardware.

Documentation

Document	What's in it
SETUP.md	Prerequisites, environment variables, running locally
ARCHITECTURE.md	System design, provider contract, resilience, dedup and ranking
ROADMAP.md	What's built, what's next, what is deliberately out of scope
KNOWN_LIMITATIONS.md	Trade-offs and gaps, written down as decisions
CHANGELOG.md	Release history
docs/specs/	Design documents, one per sub-project

KNOWN_LIMITATIONS.md is worth reading before filing a bug — several surprising behaviours are deliberate and explained there.

License

MIT © 2026 Gautam Karat