Skip to content
Thesivor
Read the docs
How it worksFeaturesYour AILimitsFAQRead the docsDownload

Local-first research workspace

A boardroom for your research idea — from first thought to a defensible draft.

Thesivor challenges your idea with AI critics, anchors every claim to its source, analyses your data and drafts in LaTeX. It runs on your computer, and your project folder is the record.

Defend before you draft.

DownloadRead the docs

Free during preview · macOS available; Windows and Linux preview · no account needed.

Thesivor's Home screen with an idea typed in, an attached PDF and note, and the starting-point chip reading “Raw idea (detected)”.
Test project, synthetic data. Start from an idea, a proposal, a dataset or an existing draft.

Why it exists

The hard part is not writing. It is knowing whether the idea holds.

An idea can survive a supervisor’s nod and still fall apart in review: the novelty was assumed, the closest prior work was never found, a number came from nowhere. Thesivor makes those attacks happen early, while they are still cheap to fix.

It does that in a fixed order — challenge the idea, gather the evidence, analyse the data, then draft — and it refuses to invent anything along the way.

Challenge first

Five critics interrogate the idea before a single sentence of the paper is written.

Evidence, not vibes

Claims are anchored to lines in real papers; numbers come from notebook cells you can point at.

Honest by default

If something has no source, it is listed as missing — not smoothed into plausible prose.

How it works

One pipeline, from a raw idea to a draft you can defend

Seven stages. You can stop at any of them — nothing forces you forward.

1Idea

Start from a raw idea, a proposal you already wrote, a dataset, or an existing draft. A folder you pick becomes the project — Thesivor never moves or edits your files.

Thesivor's Home screen with an idea typed in, an attached PDF and note, and the starting-point chip reading “Raw idea (detected)”.
Test project, synthetic data.

2Boardroom

An advocate interviews you until it understands the idea, then five critics challenge it in order. You answer, the board rules on your answer, and the round ends in a verdict.

The boardroom's advocate asking two intake questions above a ledger of eight items marked established or uncertain.
Test project, synthetic data.

3Literature vault

Search open scholarly sources, accept the papers you want, read them with every claim anchored to the exact line, then run a method that writes cited notes and a matrix — and drops any claim it cannot source.

The literature method showing cited findings with page anchors, inferences kept apart from findings, and one claim dropped by the checks.
Test project, synthetic data.

4Data & notebook

Bring a dataset or collect one. The project gets its own Python environment, downloaded on demand, and a real notebook for the analysis behind every number in the paper.

A dataset profile listing each column's type, missing values and examples, with three suggested research questions underneath.
Test project, synthetic data.

5LaTeX draft

Write in Code or Visual mode with a live PDF, a TikZ editor, and versions you can freeze. Tectonic downloads on first use; nothing about your draft leaves the project folder.

The LaTeX workspace in Visual mode: the paper drawn as a document with section chips, a file tree and a toolbar.
Test project, synthetic data.

6Proposal

Generate a research proposal, a supervisor one-pager, a first-contact email and a funding proposal — each a real document, each built only from your brief, your data and your vault.

A funding proposal built from a public NSF call, showing the extracted sections, the page limit and the review criteria.
Test project, synthetic data.

7Review

Paste the reviewers' comments, work through them one by one with a draft reply, and get a response letter with page and line locations for every change. Optionally, run a mock review before you submit.

The reviews inbox with comments split per reviewer and a confidentiality banner saying reviews are kept out of git.
Test project, synthetic data.

What you get

Every stage, in detail

Each section below says what it does for you, with a screenshot of the real app.

The boardroom: five critics before you write a word

Most ideas fail because nobody attacked them early. The boardroom makes the attack part of the work: an advocate learns your idea, then a novelty skeptic, a significance skeptic, a feasibility hawk, a soundness auditor and a publication strategist each challenge it in turn. You answer in your own words, and the board rules on whether your answer actually addresses the objection.

  • An advocate interviews you first and keeps a visible ledger of what it still does not know.
  • Five critic lenses run in a fixed order, one turn at a time — not a single blast of questions.
  • Your rebuttal is adjudicated: resolved, downgraded, or told plainly that it still stands.
  • The round ends in a verdict whose scores cite the objections that justify them.
  • Quote a message or @mention a critic to answer it directly.
  • Nothing needs a reload: every step is visible, retryable and stoppable.
Read the guide
A boardroom exchange: the Novelty Skeptic's ruling, the board reading the rebuttal as new information, and the advocate's defence.
Test project, synthetic data.

The literature vault: read papers, not snippets

The vault collects papers from open scholarly sources, lets you accept or reject each suggestion, and then makes you read them properly. Every claim you keep is anchored to the exact line it came from, so a citation in your draft can point back to the sentence that supports it.

  • Suggestions come from OpenAlex, Crossref, Semantic Scholar, arXiv, Europe PMC (including PubMed), Zenodo and OpenAIRE Graph; CORE and Unpaywall work with a key or an email.
  • Ranked candidates with a reason and the question each one addresses; only what you accept is downloaded.
  • A continuous PDF reader with selectable text and click-to-line anchoring — no OCR, so a scanned PDF is refused with a clear message.
  • A literature method that searches, fetches, writes cited notes, finds gaps and builds a matrix; a claim it cannot quote is dropped with the reason.
  • Table, figure and equation reading: detection runs locally, and chart values are always marked as estimates.
Read the guide
The literature matrix by paper, with page anchors, an “abstract only” marker, and papers listed as needing manual upload.
Test project, synthetic data.

A notebook with its own Python environment

Numbers in a paper must come from somewhere you can show. The Data & Analysis panel gives every project its own Python environment, so a package you install for one paper never breaks another.

  • Import CSV, TSV, Excel or Parquet files and get a profile of the data before you analyse it.
  • The Python runtime and base packages download on first use; heavier packs download only when a feature needs them.
  • Downloads are resumable and checked against a pinned SHA-256 before they are used.
  • Real code cells, markdown cells, charts, and outputs you can send to the LaTeX draft at natural size.
  • Nothing Python ships in the app — the installer stays small.
Read the guide
A dataset profile of a small CSV with column types, missingness and example values.
Test project, synthetic data.

A LaTeX workspace that stays out of your way

Write the paper in the editor you already know, with a live preview and the tools that are actually painful in plain LaTeX: tables, figures, math, diagrams and versions.

  • Tectonic is downloaded on first use, pinned and hash-checked; a system TeX install is detected and offered.
  • Code mode with highlighting, autocomplete, a lint gutter and SyncTeX both ways.
  • Visual mode over the same document: headings, lists, math, figures, tables, citation and reference chips.
  • A TikZ editor that lets you drag the coordinates it can handle, and tells you plainly which ones it cannot.
  • Freeze a version with its compiled PDF and SyncTeX file; restore or diff it later.
Read the guide
The TikZ editor with the picture's source on the left and a rendered diagram with draggable handles on the right.
Test project, synthetic data.

Drafting with a gate, and every number traced

AI can write plausible sentences with invented numbers. Thesivor refuses to. Drafting is blocked while assumptions are pending, and before you freeze a version the app checks that every AI sentence, number, citation and figure has a real source.

  • One section at a time, in dependency order: Methodology → Results → Related Work → Introduction → Conclusion → Abstract → Title.
  • Every drafted number must come from a named notebook cell; every citation from a paper in your vault.
  • A traceability check lists AI text with no source, numbers that are not in the notebook, and placeholders — and blocks Freeze unless you waive it with a reason.
  • Figures are inserted one at a time and are not done until you sign them off.
  • The paper checklist shows, per section, what is present, drafted, grounded and approved — and what it still needs.
Read the guide
A traceability pass showing traced spans, a figure reference, and the per-section checklist with grounded sections marked.
Test project, synthetic data.

The proposal stage: proposal, one-pager, email, funding

Before the paper, there is usually a supervisor and a call for funding. Thesivor turns the same brief and evidence into the documents those conversations need.

  • A research proposal as its own LaTeX document, with the same editor, preview and SyncTeX.
  • A supervisor one-pager and a first-contact email, with the word count done by the app.
  • A funding proposal built from a funder's call you paste or upload: extracted sections, a criteria → sections checklist, and an over-page-limit warning.
  • Discussion points only from your own open questions; risks only from the objections that survived; references built from your vault only.
Read the guide
A funding proposal built from a public NSF call, with the extracted sections, page limit and review criteria.
Test project, synthetic data.

The review phase: answer real reviewers

Peer review arrives as a wall of comments. Thesivor keeps them attached to the exact version that was reviewed, helps you work through them one at a time, and writes the response letter with a location for every change.

  • A reviews inbox per frozen version; paste or import .txt, .md, .pdf, .docx or .eml, with deterministic parsers and a manual split you can fix.
  • Reviews are confidential: kept out of git by default, never deposited, never sent to a CLI agent's workspace.
  • An optional mock review before you submit, using the venue's own form and scales, labelled “Mock (AI)” and never put in the letter.
  • Address with AI proposes a diff you apply hunk by hunk, with the same invariants as drafting.
  • A response letter that compiles, with page and line locations, and a changes PDF between the two versions.
Read the guide
The revision workbench with an “Address with AI” diff and a guard asking to confirm a change to a number.
Test project, synthetic data.

Edit for your voice — and say how AI was used

The last pass is making the paper sound like you. The voice editor learns from samples of your own writing and proposes changes you apply hunk by hunk — while protecting numbers, citations, math and meaning.

  • Paste or upload writing samples; a style profile is computed locally and you can edit it.
  • Every proposed change is checked in code before you see it: numbers, \cite keys, math, commands, comments, references, quoted text, URLs and meaning words are protected.
  • A change with no writing rule behind it is not made — no synonym spinning, no detector-score games.
  • An optional AI-use statement built from the document's own record: what the AI drafted, what you approved, which model. It is never inserted automatically and never required.
Read the guide
Edit for my voice with two proposed hunks, each with Apply and Reject, and a list of changes withheld because they would touch protected content.
Test project, synthetic data.

Deposit to Zenodo — as a draft you publish yourself

When the paper, the data or a figure is ready, Thesivor prepares a Zenodo deposit for you: metadata validated, files listed with sizes, a DOI reserved.

  • A frozen paper or proposal version, a dataset with a generated README and SHA-256s, the software, or one signed-off figure.
  • The exact file list is shown with sizes; keys, voice samples and internal record files are never uploaded.
  • The app creates a draft and stops. It has no publish action: you open Zenodo and publish it yourself.
  • The token is stored in your OS secure storage, never in the project.
Read the guide

Talk to it, and let it read to you

Typing long rebuttals is tiring. Every text box has a microphone, and every reply can be read aloud.

  • Press the mic to record; the words are appended to what you typed, and the OS microphone light goes off when you stop.
  • Local speech-to-text runs offline after a one-time model download — except on Intel Macs, where no build exists.
  • Gemini or OpenRouter can transcribe instead, if you prefer.
  • Listen, or turn on auto-read, for boardroom replies, notebook answers and editor chat.
Read the guide
The Home composer with the microphone button in the corner of the text box.
Test project, synthetic data.

Bring your own AI

Use the model you already trust — or a free one

Thesivor does not bundle a model. You connect one, and you can use a different engine for each stage.

Engines are mixed per stage and per agent: a cheap model can interview you, a strong one can draft, and a local model can read the literature. Every choice is yours; nothing is bundled.

  • OpenRouter

    One API key; the app can pick a free model for you.

    Good for: The quickest way to start, and the cheapest.

    Free models have daily limits and can be slow or fail; the app retries and offers to switch.

  • Anthropic API

    Your own API key.

    Good for: Strong long-form writing and careful reasoning.

  • Google Gemini

    A key from Google AI Studio; the app also offers Google sign-in through Antigravity (agy).

    Good for: Long context and images, including reading tables and figures.

  • A local server (Ollama, LM Studio)

    Point the app at the server's address; a key is optional.

    Good for: Keeping every prompt on your own machine.

  • Claude Code (CLI)

    Uses your existing Claude Code sign-in; the app never reads its credentials.

    Good for: Agentic editing inside a project workspace.

  • Codex (CLI) — Connect ChatGPT

    Connect ChatGPT through the Codex CLI, which opens the browser to sign in.

    Good for: Using a ChatGPT plan you already pay for.

    Sign-in is through Codex; the app never sees your password.

  • opencode (CLI)

    Uses the CLI's own sign-in.

    Good for: Open models through a command-line agent.

  • Antigravity (agy)

    Uses your Google sign-in and plan.

    Good for: Google's models without an API key.

“Connect ChatGPT” signs in through the Codex command-line tool, which opens the browser. The app never sees your password. Set up an engine.

Your work stays yours

Local-first, by design

The project folder is the record, keys live in your system's secure storage, and reviews stay confidential.

  • Your project folder is the record

    Everything Thesivor writes lives in the folder you chose. The project's record is tracked in git, so a diff shows what changed and a clone keeps the reasoning. The app's cache can be rebuilt from the record.

  • Keys stay in your system’s secure storage

    API keys go to the operating system's secure storage, never to a file, a log or the page you see. A command-line engine uses its own sign-in; Thesivor never reads its credentials.

  • Reviews stay confidential

    Peer reviews are kept out of git by default, never deposited to Zenodo, and never placed in a command-line agent's workspace. Review tasks run only on the engine you pick for them.

  • You decide what is sent where

    Only the material a step needs goes to the engine you chose for that step. PDF text is treated as untrusted data, never as instructions. Audio is held in memory and sent once unless you turn on keeping recordings.

  • No telemetry

    The app has no analytics, and neither does this site. There is no newsletter and no blog. The only network calls are to the sources, engines and downloads you ask for.

The full data map is in the privacy page. There is no log file yet; the troubleshooting page shows how to capture the app’s output from a terminal.

Get Thesivor

Small download. Everything else on demand.

The installer carries no Python and no LaTeX engine. They download the first time a feature needs them, each checked against a pinned hash.

Preview builds are unsigned.

macOS will warn you the first time, and Windows SmartScreen will say “Windows protected your PC”. The app is not damaged — the install guide shows the exact steps, including the terminal command for macOS.

macOS — available

The tested builds. Pick Apple silicon (M1 and later) or Intel.

Windows and Linux — preview, not yet tested

The installers are built, but the app has not been launched on these systems yet. Try them, but expect rough edges.

Free during preview. Nothing is uploaded by this page; there is no analytics on this site.

Settings → Storage listing the Python runtime, packages, Tectonic, the table-detection model and local speech-to-text as things provided on demand.
Test project, synthetic data. Settings → Storage shows exactly what is installed and how much space it uses.

Evidence

How Thesivor performs on a real paper

We would rather show you nothing than show you a number we cannot stand behind.

How Thesivor performs on a real paper

Results coming soon.

The test takes a published paper and replays it through the whole pipeline, then compares the result with the original — measured by code and by independent judges. We will publish the numbers here when the run is complete and checked. Until then, this page shows no figures.

Read before you install

What Thesivor does not do yet

These are real limitations in version 0.1.0, not marketing caveats. Each one links to the page that explains it.

  • The builds are unsigned

    Details

    macOS will warn you and may ask you to open the app from System Settings; Windows SmartScreen will say “Windows protected your PC”. The install guide shows the exact steps. Signing is planned, not done.

  • Windows and Linux are preview, not yet tested

    Details

    The installers exist, but only the macOS builds have been launched and checked. Treat the Windows and Linux builds as preview until the manual test has been run on real machines.

  • No OCR for scanned PDFs

    Details

    If a PDF has no text layer, the vault refuses it with a clear message. There is no OCR in this version, so you cannot read a scan of a printed paper.

  • Local speech-to-text is not available on Intel Macs

    Details

    There is no Intel-Mac build of the local speech model, so the app offers Gemini or OpenRouter transcription instead. Apple silicon Macs, Windows and Linux can run it locally.

  • Table and figure detection is imperfect and slow

    Details

    Detecting tables, figures and equations takes roughly 12–54 seconds per paper on the machine it was measured on, and reading them with a model takes longer still (about 92 seconds for a two-page paper, 278 seconds for three charts). Precision and recall vary by document type, and chart values are always estimates.

  • Zenodo deposits are drafts; the live path is untested

    Details

    The app creates a draft and never publishes it — you publish on Zenodo yourself. The whole flow has only been tested against a local mock, never against the real Zenodo server.

  • The AI-use statement is optional and never pushed

    Details

    The app can generate a short statement of how AI was used. It never inserts it on its own, never reminds you, and never requires it to export.

  • You bring your own AI engine

    Details

    No model is included. Free OpenRouter models work but have daily limits and can be slow or unavailable. Nothing that needs a model works until you configure one.

  • There is no log file or diagnostics export yet

    Details

    The app's messages go to the console; it does not write a log file. To report a problem, run the app from a terminal and copy the output — the troubleshooting page has the exact commands for each OS.

Smaller things worth knowing

  • Importing an existing single-file paper cannot be drafted section by section; the app says so instead of guessing.
  • dblp is a link-out only, and bioRxiv/medRxiv searches run through Europe PMC. DOAJ and PubMed E-utilities were not built.
  • CORE needs a key and Unpaywall needs an email address; Semantic Scholar works without a key but its free pool rate-limits.
  • Tectonic's first compile is fast only after a background package preload, and that preload needs a network connection the first time.
  • Kaggle and warehouse imports are not built.
  • There is no collaboration, cloud sync, hosted version, auto-update or mobile app in 0.1.0.

Questions

Frequently asked

Is it free?

Yes — free during preview. There is no licence statement and no announced future price. If you use a paid AI engine, that engine bills you directly; the app itself is free.

Do I need an API key to use it?

Only for the steps that call a model. You can install the app, open a project, read papers and compile LaTeX without one. Free OpenRouter models work for the model steps, with daily limits; a local server such as Ollama needs no key at all.

Read more
Does it work offline?

The app is local-first, and the notebook, the PDF reader and LaTeX compile offline once their tools have been downloaded. Model steps need the engine you chose, so a cloud engine needs a connection. A local engine can work fully offline.

Where does my work live?

In the project folder you choose. The project's record is tracked in git inside that folder; the app's cache can be rebuilt from it. Keys are stored in your operating system's secure storage, not in the project. Reviews are kept out of git by default.

Read more
Is the paper written by AI?

No. The drafter can propose a section, but every sentence it writes must trace to your brief, your data or a paper in your vault, and you approve each section. Before you freeze a version, the traceability check lists anything without a source. The AI-use statement is optional and describes what actually happened.

Read more
Why does macOS say the app is damaged or from an unidentified developer?

Because the preview build is unsigned. The install guide shows the exact steps to open it, including the terminal command for the “damaged” message. It is not actually damaged.

Read more
Does it work on Windows or Linux?

The installers are built, but the app has not been launched or tested on Windows or Linux yet. Treat those builds as preview. The macOS builds are the tested ones.

Read more
Do I need Python or LaTeX installed?

No. The app downloads its own Python runtime and its own LaTeX engine the first time you need them, each checked against a pinned hash. If you already have a system TeX install, the app detects it and offers it.

Read more
Can I use a ChatGPT subscription I already pay for?

Yes, through Connect ChatGPT, which signs in through the Codex CLI. The app never sees your password, and it uses the CLI's own sign-in.

Read more
Is this a replacement for my supervisor?

No. It is a boardroom, not a person: it challenges the idea and keeps the evidence in order. The proposal stage even generates a one-pager and a first-contact email so the conversation with a real supervisor starts from something concrete.

What happens to my reviews — are they uploaded anywhere?

Reviews are confidential. They are kept out of git by default, never deposited to Zenodo, and never placed in a CLI agent's workspace. Review tasks run only on the engine you explicitly pick for them.

Read more
Can I run it on a modest laptop?

The installer is small because Python and LaTeX download on demand. Local table detection and local speech-to-text are the heaviest parts; you can turn table detection off per paper and use a cloud engine for speech. There is no stated minimum specification yet.

About

Built by a researcher, for researchers

Thesivor is built by Akhlak Hossain (akhlak.dev). The premise is simple: an AI can write a plausible paper, but plausible is not defensible. So the app spends its effort on the part that is hard to fake — challenging the idea, keeping the sources, and tracing every claim.

It is free during preview. You can reach the author at akhlak@thesivor.akhlak.dev.

The app's Settings → Appearance section showing the theme options, rendered in the dark theme.
Test project, synthetic data. Light and dark themes, following your system by default.

Defend before you draft.

Download the preview, or read the documentation first.

DownloadRead the docs