Sciwand is a desktop AI assistant for your document library. It lets you chat with PDFs, Word files, EPUB books, and plain text (papers, contracts, books, reports, transcripts), search academic databases such as Google Scholar, PubMed, arXiv, and Semantic Scholar, and write grounded in your own sources so the AI does not hallucinate references. Aimed at researchers, students, lawyers, analysts, and anyone making sense of a stack of documents.

Which reference managers does Sciwand connect to?

Sciwand syncs with Zotero, EndNote, Mendeley and Bookends libraries, with more reference managers on the way. The sync is one-way, so your existing library is never modified.

Does Sciwand work offline with local AI models?

Yes. Sciwand supports local LLMs in addition to cloud providers, so you can work fully offline and keep your text away from any third-party provider.

Sciwand is currently in free public alpha for macOS, Windows, and Linux.

How does Sciwand avoid hallucinated citations?

Sciwand uses retrieval-augmented generation (RAG) over your own library and academic databases. Every answer is referenced so each claim can be traced back to its source PDF.

The reference and knowledge manager with an academic writer built in.

Search, screen and cite with AI, on the library you already have.

Download for free See what it does

macOS
Windows
Linux

Sleep and learning

24 papers Add AI column

Paper Relevance Effect? Key finding

Sleep and exam performanceOkano et al. · 2019 9 Yes Longer sleep in the week before an exam predicted higher grades.

Sleep and memory consolidationWalker & Stickgold · 2006 8 No

Napping and recallLovato & Lack · 2010 8 Yes

Spacing versus crammingCepeda et al. · 2015 7 Yes

Chronotype and morning examsPreckel et al. · 2013 6 Yes

Screens before bedHale & Guan · 2015 2 No

Relevance to my topic

9/10

Exam-week sleep measured against grades: directly on topic.

From the paper

Students who slept longer and more consistently in the week before the exam earned higher grades.

Okano et al. 2019 · page 4, highlighted

Connects to the library you already have

Zotero
EndNote
Mendeley
Bookends

No repeat importing. Your library stays exactly as you left it.

01Search

Ten online databases. Screened in one sitting.

One query goes to every database at once and the results merge as they arrive. The same AI columns you use on your own papers then screen the whole list.

Ten databases at oncePubMed, Google Scholar, arXiv, Semantic Scholar and six more, merged as they arrive.
Screen the whole listA score column and a yes/no column read every result for you.
Star the keepersFavourite the high scorers, fetch their full text, and file them in a collection.

02Screen and analyse

Ask one question of three hundred papers.

Describe what you want out of every paper. The model designs the columns, reads each PDF, and fills the grid, with every cell cited to a sentence.

Sleep and learning

24 papers Add AI column

Paper Population Effect size Evidence

Sleep and exam performanceOkano et al. · 2019 Undergraduates, n = 214 d = 0.42 8

Sleep and memory consolidationWalker & Stickgold · 2006 Review, 92 studies not reported 6

Napping and recallLovato & Lack · 2010 Meta-analysis, 41 trials g = 0.31 9

Screens before bedHale & Guan · 2015 Adolescents, n = 1,486 r = .18 5

Caffeine and next-day recallSmith et al. · 2020 Adults, n = 88 d = 0.24 4

Spacing versus crammingCepeda et al. · 2015 Undergraduates, n = 176 d = 0.61 8

One brief, several columns

Who was studied, how big the effect was, and how strong the evidence is.

Population Text
Effect size Text
Strength of evidence Score 1-10

Effect size · Okano et al.

d = 0.42

Students who slept longer in the exam week scored 0.42 SD higher on the final.

Page 5 · opens on the sentence

Describe it onceThe model proposes the set of columns and their types; you approve before anything runs.
Every cell has a sourceClick a value and the PDF opens on the sentence it came from.
Score, sort, filterScores and yes/no answers turn a pile of papers into a ranked shortlist.

Datasets

One paper, many rows.

Say what one row is, such as each experimental condition, and every paper contributes as many rows as it describes.

The library cell holds the count; the rows form their own table.
Fields you add later fill in without losing your edits.
Export straight to a spreadsheet for analysis.

Conditions

Each experimental condition 13 rows

# Paper Condition n Outcome Effect

1 Okano et al. 2019 Short sleep n = 71 Exam grade d = 0.42

2 Typical sleep n = 74 Exam grade reference

3 Extended sleep n = 69 Exam grade d = 0.31

4 Walker & Stickgold 2006 Sleep n = 24 Recall at 24 h d = 0.55

5 Wake n = 24 Recall at 24 h reference

6 Lovato & Lack 2010 Nap, 10 min n = 30 Alertness d = 0.38

7 Nap, 30 min n = 30 Alertness d = 0.44

Dataset column

Each experimental condition

Condition Text
n Number
Outcome Text
Effect Text

Run on all 24 papers

13 rows from 5 papers

Prompt types

Six ways to ask a question of every paper.

Pick the shape of the answer, and the column fills itself across the whole collection. Start from a suggested prompt, or one you saved.

PaperKey finding

Longer sleep predicted higher grades.

Sleep after learning stores it.

Naps recover part of recall.

Text

The model reads each paper and writes a short passage in the cell.

PaperRelevance

Score

A number from 1 to 10 against criteria you set, drawn as a bar so the column sorts and scans.

PaperRandomised?

Yes

Yes/No

A yes or no for every paper, with the reasoning kept behind it.

PaperPopulationEffect

n = 214d = 0.42

92 studiesnone

41 trialsg = 0.31

Multi-column

One brief, several ordinary columns. The model designs the set and picks each one's shape, and you review it before anything is created.

PaperConditions

3 conditions

2 conditions

4 conditions

Dataset

One paper, many rows. Say what one row is, a patient in a case series or an arm of a trial, and every paper contributes as many rows as it describes.

PaperMy notes

Check n = 88 subgroup

Cite in chapter 2

Plain

An empty column you type into yourself. No model runs and nothing is generated.

03Library chat

Ask your whole library.

One question, put to every paper in a collection at once. The answer comes back from your own documents, never from model memory, and every claim carries its source.

Library chat

Does a nap recover what a short night costs?

Read 34 papers

Yes, in part. A 60 to 90 minute nap restores alertness and recovers a measurable share of what the short night cost1, and the slow-wave sleep inside the nap predicts how much comes back2. Naps that reach only stage 2 lift alertness without a recall benefit3.

Lovato & Lack, 2010 Walker & Stickgold, 2006 Diekelmann & Born, 2010 Okano et al., 2019

Lovato & Lack, 2010page 3 of 14

The effects of napping on cognitive functioning

Progress in Brain Research (2010)

Naps of 60 to 90 minutes recovered 40 to 60 per cent of the recall deficit.

Citation 1

Naps of 60 to 90 minutes recovered 40 to 60 per cent of the recall deficit.

Lovato & Lack, 2010 · page 3

Opens on the page

Click a citation and the PDF opens beside the answer, the sentence already highlighted.

Cited to a sentenceEach claim carries the passage it came from, not just the paper.
Check it in one clickThe source opens at the right page so you can read the context yourself.
Nothing inventedIf a claim cannot be anchored to a passage you own, you see that instead of a made-up reference.

04Write

A next-generation academic writer.

Cite at the caret in any of ten thousand styles, pin each citation to the sentences it stands on, and make figures from the numbers already in your paper. Or do all of it inside Microsoft Word.

Writer

Neurotransmission at the chemical synapse

Background

Signalling between neurons is chemical, not electrical, across the great majority of connections in the mammalian brain [1]. The delay this introduces is the price of a junction that can be modulated rather than merely conducted [2].

Objective

We set out to measure how the readily releasable pool of vesicles recovers after sustained stimulation [3]. Every recording was scored blind to the behavioural data [4].

Figure 1. Pool depletion by stimulation frequency.

10,000+citation styles

APA 7th
IEEE
Nature
Vancouver
Chicago

Switch, and the list and the markers follow.

[4] Okano et al., 2019

Pinned to this sentence, not the whole document.

123

Quote 2 · PDF page 4

New figure

5 Hz 20 Hz 50 Hz EGTA

horizontal bars, a colour per conditionInsert

Any journal's style10,000+ citation styles; references reformat themselves as you switch.
Cited per sentenceEach marker is pinned to the sentence it supports, with the quoted passage behind it.
Figures without leavingAsk for a change in words and the chart redraws, then drops into the page.

05Microsoft Word add-in Beta

Cite from your library without leaving Word.

A SCIWAND tab in Word's own ribbon, on Windows and macOS. Search your library, insert real citations in any style, and keep the bibliography in step, inside the document you are already writing.

A tab in Word's ribbonAdd and edit citations and the bibliography from a SCIWAND tab, on Windows and macOS.
Citations that keep upThey update and restyle as you edit, renumbering in numbered styles, and the reference list stays in step.
The same AI, in your draftAsk your library, ask the document in front of you, or improve a selection.

06Read and connect

Every paper, and everything around it.

A paper opens with its PDFs, notes and highlights in place. Its references and the work that cites it are drawn as a map, so the shape of a field is visible at once.

PhD work

Sleep, learning and the developing brainKurdziel et al. · 2013

Slow-wave activity and memoryMarshall et al. · 2006

Sleep, consistency and gradesOkano et al. · 2019

The memory function of sleepDiekelmann & Born · 2010

Distributed practice in recallCepeda et al. · 2006

Napping and recallLovato & Lack · 2010

Journal article

Sleep duration, bedtime consistency and academic performance

InfoFiles 5Refs 8Cited by 5Graph

Cited byReferencesRelatedShared

Note · Okano 2019

Consistency predicted grades beyond total sleep. Check the n = 88 subgroup before citing.

Torn off into its own window.

Kurdziel 2013

Sleep, learning and the developing brain

Shares 12 references

Notes beside the paperPull a note out as its own window and keep it open while you read.
The citation mapReferences, citing work and related papers, colour-coded and clickable.
A real readerPDF and EPUB, highlights and annotations, all stored on your own disk.

The full picture

Everything else in the box.

Ten languages

The whole interface in ten languages, including Arabic and Persian, laid out right to left.

Annotate anything

PDF, Word, EPUB, Markdown, plain text and saved web pages, with highlights and notes.

Tables out of PDFs

Extract a table from a paper and turn it into rows you can score and compare.

Charts and diagrams

Charts from a spreadsheet or a pasted table, and scientific diagrams drawn to order.

Your data, your call

Everything lives on your disk. Cloud sync is optional, category by category.

Local models

Bring your own API key, or run a model on your own machine and let nothing leave it.

Fits where you already are

Nothing to migrate. Nothing to give up.

Sciwand works standalone or beside the tools you have. Imports are one-way, so your reference manager stays exactly as you left it.

Zotero
Mendeley
EndNote
Bookends
Microsoft Word
Obsidian
Bear
Notion

Any document you have

Drop a file, a whole folder, or paste a DOI, PMID, arXiv ID, ISBN or a title.

One-way sync

Nothing is written back to Zotero, EndNote, Mendeley or Bookends. Your library is never modified.

Your model, your key

Use the provider you already pay, or a local model.

How it compares

Where Sciwand differs.

Zotero, EndNote and Mendeley keep your references in order. Sciwand does that too, then lets you search, analyse and write with them.

Based on each product's public documentation. Spot something out of date? Tell us at [email protected].
Capability	Sciwand	Zotero	EndNote	Mendeley
Stores your references, PDFs and metadata	Yes	Yes	Yes	Yes
Cites in Microsoft Word, renumbering as you edit	Yes	Yes	Yes	Yes
Thousands of citation styles	Yes	Yes	Yes	Yes
Ask a question across your whole library, answered with citations	Yes	No	No	No
Every answer traced to the sentence it came from	Yes	No	No	No
Screens hundreds of PDFs into a comparison grid	Yes	No	No	No
A document writer inside the app, citation engine underneath	Yes	Notes only	No	No
Citations that carry the quoted passage, not just the paper	Yes	No	No	No
Imports your existing library, one way, nothing to migrate	Yes	Partial	Partial	Partial

Questions

The things people ask first.

Is this a reference manager or a document writer?

Both, and that is the point. One side is a full reference manager: collections, metadata, PDFs, annotations, notes, a citation graph, and sync with Zotero, EndNote, Mendeley or Bookends. The other is a real writing environment with its own editor, or the Word add-in if you would rather stay in Word. A paper you are reading is one keystroke from being cited in the paragraph you are writing.

Is it free?

Everything except the AI is free, permanently. No account and no time limit: the whole reference manager, PDF and EPUB reading, annotations and notes, the writer, the Word add-in, the citation engine with its 10,000+ styles, per-passage quote linking, and sync with Zotero, EndNote, Mendeley and Bookends.

The AI features come with free one-time credit to get started. Past that you need one of these, never both.

Sciwand AIA subscription, and only for the AI. We run the models, so there is nothing to configure. It is also your licence while it is active.
A licenceOne off, from $29.99, lifetime updates. Bring your own API key and pay your provider at cost, or run a local model and pay nothing at all.

Will it hallucinate citations?

That is the failure mode Sciwand is built around. Answers are generated against the documents in your library rather than from model memory, and each one carries a badge that opens the source PDF at the page with the sentence highlighted. If a claim cannot be anchored to a passage you own, you will see that.

Does my work leave my computer?

Only if you ask it to. Your library, annotations and notes are stored on your own disk, and cloud sync is optional: you choose which categories sync, if any. With a hosted model, the text needed to answer a question goes to the provider whose key you supplied, and nowhere else. Run a local model and nothing goes to a provider at all.

Do I have to abandon Zotero or EndNote?

No. Sciwand syncs with Zotero, EndNote, Mendeley and Bookends. The sync is one-way, so nothing is written back and your existing library is never modified. Plenty of people run it purely as a reading and analysis layer on top of what they already have.

Does it work with Microsoft Word?

Yes. The add-in puts a SCIWAND tab in Word's ribbon on Windows and macOS, with your library, real citations in whichever style you need, and the same AI you get in the app. It is in beta, and it installs from the app's own settings.

Which databases can it search?

Ten academic databases from inside the app, including PubMed, Google Scholar, arXiv and Semantic Scholar. Results can be screened with the same AI columns you use on your own documents, so discovery and screening are a single pass.

Get started

Point it at your own sources and see.

On macOS, Windows and Linux, with the Word add-in for Windows and macOS. Your library stays on your machine.

Download for MacUniversal · Apple Silicon + Intel Download for Windowsx64 installer Download for LinuxAppImage · x86_64

Latest release · ARM64 and all builds · read the docs