IncarnaMind: page-level citations for AI writing on your own documents

A notebook built around files you already have

IncarnaMind is a free, open-source desktop app for macOS, Windows and Linux that is written in TypeScript and packaged as an Electron program. The repository root carries the usual Electron markers, among them electron-builder.yml and electron.vite.config.ts, and the licence is Apache-2.0.

The problem it targets is narrow and real. Plenty of assistants will summarise a folder of papers or a pile of contracts, but they rarely tell you where each statement came from, so the work of checking moves back to you. IncarnaMind builds the citation into the answer instead. Every answer carries the quote it relied on and the page it sits on, and the quote is itself checked against that page.

The people this fits are researchers working from a reading list, students comparing several papers on one topic, and analysts who live inside PDFs and spreadsheets and need every figure traceable. The unit of work is called a Mind, a notebook for one topic, and beside it sits an assistant you can question about that Mind or about your files.

Check against Documents returns a verdict per sentence

The feature that separates IncarnaMind from a note pad with a chat panel is a sentence level check. You select sentences in a Mind and choose Check against Documents. Each selected sentence comes back labelled supported, unsupported or contradicted by your Documents, and the page that settles it is one click away. A badge in the corner of the page counts how much still needs a look, so a long draft shows its unresolved claims before you send it.

The same check applies to work you did not write. For a PDF someone else produced, the project describes a path called Verify a document, so you can work through a claim by claim instead of reading the whole file.

This is a different contract from the usual one. An assistant that answers with a paragraph leaves you to spot which clause came from where. IncarnaMind instead makes every sentence carry a status, which means the failure mode is visible as a list of amber marks rather than hidden inside fluent prose. Note what it does not claim: it does not state that the supported label is a judgement about whether the claim is true, only that the Documents support it.

Quotes get checked a second time against the page

An answer in IncarnaMind cites where each claim came from and carries the quote from that place. That quote is then verified: a green check when it is actually present, an amber mark when it is not, and a click opens the page with the quote highlighted so you can see it in context.

Unfolding an Answer shows what it searched and which sources it used, which answers a question most assistants ignore, namely which chunk of your corpus the answer actually came from.

Writing happens in the same panel. Your draft stays exactly as you wrote it until you pick Add to note, which places the answer at your cursor as text you can edit, citations included. The assistant can also be asked to chart data, for instance gross margin by quarter, which produces a cited table and a bar chart attached to the Mind. Model choice is deliberately open: your ChatGPT, Claude or Copilot plan, or your own API key.

Meeting transcripts stay on the machine that recorded them

IncarnaMind includes a meeting mode. You start recording and a live transcript runs beside your Mind, transcribed on this computer. When the call ends you can save the transcript as a Document, which means later notes can cite what was said rather than paraphrasing it from memory.

During the call, a small panel can suggest what to say and answer a question from your Documents. The app also displays a notice telling the other participants that the call is being recorded, which is the kind of detail that matters more in practice than any feature list.

The privacy claim here is specific and limited: the transcription runs locally. That is a narrower statement than saying nothing leaves your computer, since answering your questions can still involve the model you have configured. The project describes local-first in its topic list and states that linked folders are read where they are and never moved or changed, but the documentation does not spell out an offline mode with a bundled model.

Linked folders and citation sources that arrive without typing

Instead of uploading copies, you link the folders you already use, and the project names Zotero's storage folder as an example. IncarnaMind reads PDF, Word including comments, PowerPoint, Excel, CSV, Markdown and plain-text files in those folders where they are, never moves or changes them, and keeps up as they change.

Sources can also arrive from outside. You can search OpenAlex, arXiv and PubMed and add open-access papers as Documents, save a web page by its address so that it can be cited and checked like a file, and import references from Zotero or any CSL JSON, BibTeX or RIS file.

Organisation inside the app is by Folder, Linked folder and Tag, with a workspace shown as six Folders, a Linked folder and Tags. An Organize action can read a folder of new Documents and assign each one a Suggested Folder and Tags. The practical consequence is that the citation trail depends on file formats the project can parse, and the formats listed above are the ones it names.

There is no installer yet, so the first run is from source

The project has not shipped a release. The download section states that installers for macOS on both Apple silicon and Intel, Windows and Linux will appear on the Releases page, and that the current way in is to run it from source with Node.js 24 or newer. That command is the whole setup:

git clone https://github.com/IncarnaMind/IncarnaMind.git && cd IncarnaMind
npm install && npm run dev

The engine constraint is recorded in the manifest, along with the build and preview entry points:

  "engines": {
    "node": ">=24"
  },
  "scripts": {
    "dev": "electron-vite dev",
    "build": "electron-vite build",
    "start": "electron-vite preview",

Linux packaging is described as a .deb for Ubuntu and Debian, recommended, or an AppImage. The AppImage needs the execute bit set, and on Ubuntu 22.04 or later a fuse package:

chmod +x IncarnaMind-*.AppImage

sudo apt install libfuse2

macOS needs a quarantine exception until the app is notarized:

xattr -dr com.apple.quarantine /Applications/IncarnaMind.app

On Windows the installer is unsigned, so SmartScreen blocks it until signing through the SignPath Foundation arrives. On macOS an unsigned app cannot update itself, so the app reports that a new version exists and offers its download page.

The topic list promises integrations the documentation never describes

The repository topics mention mcp, ollama, local-llm and notebooklm, and the tree contains AGENTS.md, CLAUDE.md, CONTEXT.md, an eval directory, a wasm directory and an e2e directory. None of the documentation shown describes how to point IncarnaMind at a local Ollama model, how to expose it over MCP, or how to import a NotebookLM export. Those names are tags and directories, not instructions, and treating them as features would be reading more into the listing than it supports.

The maintenance picture is healthy but young. The last push was on 2026-10-22, and three tagged pre-releases landed within a week: v0.1.0-alpha.1 on 2026-10-13, v0.1.0-alpha.2 on 2026-10-16 and v0.1.0-alpha.3 on 2026-10-18. The version in package.json is 0.1.0-alpha.3.

Early pre-release numbering plus a daily push cadence usually means breaking changes to configuration and UI. The scripts are named for that reality, with separate evaluation suites for word, excel, ocr, transcription, organize and references work. Upgrade cost for a user is therefore low but the target moves.

Apache-2.0 and the real cost of running an alpha locally

Licensing is Apache-2.0, stated in package.json and shown in the licence badge, with the LICENSE file at the repository root. That is a permissive grant with an explicit patent clause, and it is the reason a commercial team can look at the code without a procurement conversation. This is a note on the licence text, not legal advice.

The larger cost is operational. Because there is no signed release, adoption means compiling and running the program yourself on a machine with Node.js 24, then dealing with platform friction that a normal installer would hide: SmartScreen on Windows, the quarantine flag on macOS, fuse on Linux, and a manual update check because an unsigned macOS build cannot update itself.

The verification workload also lands on you. A sentence marked supported is only as good as the parser that read the PDF, and Word comments are listed as read, so documents whose meaning lives in those comments are in scope for checking. Nobody has published timing or accuracy figures for any of this, so any performance claim you hear about the project is a claim about something other than this code.

Editorial conclusion

IncarnaMind suits researchers, students and analysts who have to defend every sentence against a source file and who would rather check a draft than rewrite it. It is a poor fit for anyone who needs a stable installer today, since the project has not shipped a release and its builds are unsigned, or for teams that require an offline path without testing a local model setup themselves. Before adopting it, verify three things on your own machine: that npm install and npm run dev finish under Node.js 24, that the citation quote highlighting works on the PDF types you actually use, and that the model subscription you intend to point it at returns citations you can audit.

Frequently asked questions

Official sources

1. IncarnaMind/IncarnaMind on GitHub
2. Issues
3. License: Apache-2.0
4. README
5. Releases