Research with an agent.
Every citation verified.

abstract reads your papers, screens literature PRISMA-style, chains citations, and writes documents where every factual sentence links to the passage that proves it — with the quote on hover. Local-first: your library never leaves your machine.

Get started free Pricing
terminal
bun install -g abstract   # or: npm i -g abstract
abstract ~/my-thesis      # opens your workspace
2-minute demo watch abstract build a verified literature review
verified writing

Sentences that carry their evidence

Drafts cite only passages the agent actually opened. Every claim is checked against its source and marked supported, partial, or flagged — and the export refuses to ship fabrications. Hover any sentence to see the verbatim quote behind it.

Bibliographies render from registry metadata only — DOI-checked at export, retractions flagged. Markdown, HTML, LaTeX, and BibTeX out of the box.

Deep networks can memorize random labels entirely [1]. Regularization improves generalization only marginally in this regime [1]. Later work extends the result to modern vision transformers [2].
evidence · zhang2017.pdf · p.3 · peer-reviewed “…networks are rich enough to memorize entirely random labelings of the training data…”
screening funnel · live in the app
identified
555
screened
128
included
19
read
17
every decision logged with its reason → exportable PRISMA flow
real discovery science

PRISMA screening and citation snowballing, built in

The agent searches scholar indexes concept-first, screens candidates with recorded include/exclude reasons, and snowballs references backward and forward until saturation — the whole trail travels with your review, ready for a methods section.

Downloads are deduplicated by DOI, file, and title; metadata resolves against Crossref and OpenAlex — never from model memory.

how it works

From question to verified document

Ask

“Do a literature review on frugal AI × FATE benchmarks.” The agent plans, searches, and screens — you steer at any point.

Read

It ingests PDFs, reads page by page, takes notes, maps concepts, and builds an evidence base it can cite.

Write

Drafts are grounded in retrieved passages, claim-checked sentence by sentence, tables cell by cell.

Export

Print-grade HTML with citations deep-linking into the PDF at the page, plus Markdown, LaTeX, and BibTeX.

local-first

Your library stays yours

abstract is a CLI that opens a workspace in your browser. PDFs, notes, drafts, and the reading log live in ordinary files on your disk — not on our servers. Bring your own model keys, or run on abstract credits with per-turn receipts showing exactly what each task cost.

Even if your subscription lapses, reading your own workspace is never locked.

10-second loop
everything in the box

Built for real research, end to end

integrityVerified citations

Every sentence checked against a real passage; hover shows the verbatim quote, the page, and the source grade.

discoveryCitation snowballing

Backward and forward chaining (Wohlin-style) with saturation tracking — the agent knows when the literature is exhausted.

discoveryPRISMA screening

Search log, include/exclude ledger with reasons, live funnel, exportable flow diagram.

writingVerified tables

Extraction matrices and comparison tables where every data cell is cite-checked like a sentence.

writingGenre-aware structure

Reviews, theses, rebuttals, roadmaps, résumés — the document's shape follows its genre, never a template.

writingScholarly citation styles

Numeric or author–year, author-prominent prose ("Hooker et al. showed…"), registry-rendered bibliographies.

readingDeep reading artifacts

Print-grade HTML where each reference deep-links into your PDF at the cited page. Save to PDF from the browser.

readingFigures and visuals

The agent views pages, crops figures, and keeps what it saw pinned to its reasoning.

scaleSub-agent reading waves

Parallel focused readers with the same integrity gates — big corpora without losing the thread.

craftSkills

Venue rubrics, house styles, and methods checklists the agent loads exactly when they're relevant.

controlPlans you can steer

A visible task plan, interruptible mid-run; the agent is accountable to its own promises.

billingCredits or your keys

Managed mode with per-turn receipts (model, tokens, cost) — or bring your own API keys on the $5 license.

pricing

Start free. Stay cheap.

trial
Free

4 days, all features, 300 credits included.

byok license
$5/mo

Your own API keys, up to 3 devices. $48/yr.

pro · managed
$24/mo

No API account needed — 2,600 credits monthly, top up anytime.

Compare plans

Your next review, with receipts.

Install in one line. Free for 4 days — no card.

Get started