Finds your documents by itself
Crow counts the documents in your home folder and shows where they live: Documents, Desktop, Downloads… You tick the folders (or a date range, or single files) and see the cost and the time before anything is spent.
Crow 0.1.4 Connect Claude, ChatGPT and Codex to your wiki
Build your second brain
Build your second brain
in minutes
Turn your documents into a living wiki you can chat with.
Crow finds your PDFs, Word files, slides and spreadsheets by itself, files them into a wiki of linked Markdown notes, and answers your questions citing the notes it used.
Apple silicon · Crow 0.1.4 · free and open source · macOS, Windows and Linux
Crow's window: a question typed into a new chat, the answer comparing two kitchen quotes with the notes it cites, and those notes lighting up in the brain, the graph of the wiki beside the chat
Try it here
Explore a brain
This is Crow's brain, drawn by the app's own code. It holds the example wiki that comes with llm-wiki-v2 (llmwiki2 demo): the notes of Lumen Coffee Roasters, a small roastery in Turin that does not exist.
Drag the dots, open a folder, read a note, find one, or play a question and watch the notes it read light up.
Loading the brain…
Drag a dot · Click the brain, then scroll to zoom · Click a folder to focus it
The answers were recorded on 4 November 2026 from llmwiki2 0.8.0 in CROW mode, Jev 1.13 deciding and Gemini 3.8 Flash writing. Nothing on this page calls a model. The example wiki on GitHub.
What Crow does
A free desktop app that turns the documents already on your computer into a second brain
A wiki of Markdown notes, filed by topic and linked to each other, that you can chat with. It is the LLM wiki idea Andrej Karpathy described, built for thousands of documents and for people who would rather not use a terminal.
Files them into a wiki that organizes itself
Each document becomes a note in a folder about its subject. A new document on a subject updates the note already there instead of adding a copy, and a folder that grows is split into subfolders. Filing goes oldest first, so the newest facts win; files already filed, and duplicates, cost nothing again.
Lets you chat with your documents, with sources
Answers cite the notes they come from, and those notes light up in the brain, the graph of everything you filed. Follow-up questions are understood in context.
Remembers
A chat forgets when it ends; the wiki doesn't. Every document you add makes the notes you already have better, and you can drop a new file onto a folder of the brain to file it there.
Keeps it in plain Markdown you own
The wiki is a folder of Markdown files on your computer, in Google's Open Knowledge Format: open it in Obsidian or any editor, back it up, keep it in git. Your documents are never changed or moved.
Works with any model, within a spending limit
Gemini, Claude, DeepSeek or any model on OpenRouter, through your own account. Crow pauses at your spending limit and continues where it stopped after a restart.
Is the memory of your AI apps
One button connects Claude, Claude Code, ChatGPT or Codex: they look things up in your wiki, with its notes as sources, and keep what they learn in it. How it works
Free, and open source
Crow is free under the GNU AGPL v3.0 or later. You pay OpenRouter for the models it uses, and Crow estimates that cost before it starts.
Crow reads PDF, Word (.docx), PowerPoint (.pptx), Excel (.xlsx), CSV, OpenDocument text and spreadsheets (.odt, .ods), Markdown and plain text; a spreadsheet becomes a note per sheet. Scanned PDFs without a text layer are skipped: Crow does no OCR yet.
See it
The brain grows as Crow files
Recorded in the app, on a sample wiki made up for the recordings.
The brain growing: notes appear next to the folders of their subjects and the graph widens as more documents are filed
Zooming into the Finance folder of a large brain: its subfolders appear with their names, then a click on Receipts dims everything else
Crow's first run: the welcome screen, the documents it found by folder with a date filter, connecting OpenRouter, the cost estimate, and the wiki starting to build
Since Crow 0.1.4
Crow is the memory of your AI apps
Press Connect AI apps in Crow's sidebar, then Connect next to the app. From then on the app can ask your wiki, read a note and keep what it learns: a decision, a fix that worked, a preference, filed by Crow like any other note.
Claude the desktop app
Crow adds itself to Claude's MCP servers; on macOS, if Claude is open, the button reads Connect and restart and Crow reopens Claude for you.
Elsewhere, quit Claude first, connect, then open Claude.
Claude Code
Crow runs claude mcp add for you, for every project.
Then start a new session.
ChatGPT the desktop app
Crow adds itself to Codex's settings, which the ChatGPT app reads; on macOS it restarts ChatGPT if it is open.
Elsewhere, quit ChatGPT and open it again.
Codex CLI, IDE extension
Crow adds itself to ~/.codex/config.toml.
Then start a new session.
recallAsks your wiki. The answer cites its notes, as Crow's own chat does.
read_noteReads one note of the wiki in full.
rememberKeeps what the app learns: Crow files it, merged and linked like any other note.
It works on macOS, Windows and Linux, through MCP, on your computer only: the app runs Crow itself, which hands everything to the Crow that is running. Crow need not be open: it opens in the background the first time an app needs it.
If Crow cannot connect an app, the window says why and shows the command or the lines to paste, with a Copy button. ChatGPT Classic and ChatGPT on the web reach only servers on the internet, so they cannot use Crow: the new ChatGPT app can.
Any other MCP client (Cursor, VS Code…) can run the same local (stdio) server: Crow's own executable with the argument mcp.
Crow as a stdio MCP server
# macOS /Applications/Crow.app/Contents/MacOS/crow mcp # Windows, in the folder Crow was installed in (usually %LOCALAPPDATA%\Crow) crow.exe mcp # Linux: crow mcp, or your AppImage's path then mcp crow mcp
What a connected app recalls goes to that app's provider (Anthropic or OpenAI) as part of your conversation, under that app's terms. What it asks Crow, and asks it to remember, goes to your OpenRouter model, at the same prices as your own questions.
Why a wiki and not RAG?
Crow writes the knowledge down instead of searching it again at every question
Most "chat with your PDFs" tools use RAG: they cut documents into chunks, keep them in a vector database, and look them up again at every question. Crow reads each document once, summarizes it into a note in the right folder, merges it with what the wiki already knows and links it to related notes. Answers read those notes and cite them, and you can open, correct and keep every one.
| Crow (LLM wiki) | Typical "chat with PDF" tool (RAG) | |
|---|---|---|
| Finding your documents | Crow finds them on your computer | You upload them |
| What you keep | Readable Markdown notes, filed and linked | A vector index no one reads |
| A new document on the same subject | Updates the note already there | Adds more chunks |
| Where it lives | A folder on your computer | Often a cloud service |
CROW
A small classifier decides where things go, so the wiki stays fast and cheap as it grows
Karpathy's LLM wiki finds pages by having the model read an index, which works up to about a hundred sources. CROW makes it scale: a small classifier picks each document's folder among every folder of the wiki in one choice, and answering walks the folders level by level, so the wiki keeps growing without a vector database. The LLM writes: it summarizes each source, rewrites a note it merges into, names new folders and writes the answer. It also lays out the folders, in a plan before a batch is filed and by splitting a folder that has grown, and it decides when the classifier fails.
In the CROW paper, retrieval kept its accuracy across a twentyfold growth of the wiki.
summarize→route→find match→merge or write→relate→link, index, log
| Step | Classic | CROW |
|---|---|---|
| Summarize the source into a note | LLM | LLM |
| Route: best folder, or a new one | LLM | Classifier |
| Same subject as an existing note? | LLM | Classifier |
| Merge into it, or write a new note? | LLM | Classifier |
| Name a new folder, rewrite a merged note | LLM | LLM |
| Lay out the folders (a batch's plan, a folder's split) | LLM | LLM |
| Related notes (See also) | LLM | Classifier |
| Find the notes for a question | LLM | Classifier |
| Write the answer | LLM | LLM |
In CROW mode the LLM also decides when the classifier fails or is unsure. From llm-wiki-v2's README.
89%
right note read
CROW89%
LLM-only wiki62%
2.1×
faster per ingest
CROW18.4 s
LLM-only wiki38.6 s
68%
fewer LLM tokens per ingest
CROW5,508
LLM-only wiki17,324
These are the authors' own measurements, from the benchmark in the CROW paper (Table 2): one corpus of 525 personal documents, mostly in Italian, with one LLM and one classifier version, as the paper says among its limitations. That corpus is not public, since it holds the first author's personal notes. The corpora and scripts behind llm-wiki-v2's own measurements are public in benchmarks/, and every script has a --dry-run, so you can run them with your own models.
Download
Download Crow 0.1.4
Released 28 October 2026, for macOS, Windows and Linux. Crow is free; you pay OpenRouter for the models it uses.
macOS
This computerInstall Crow (the builds are not signed yet)
Crow's installers are not signed with an Apple or Microsoft developer certificate yet, so macOS and Windows warn you the first time you open it. That is expected for an unsigned app; here is how to open it. You do this once.
macOS "Crow" Not Opened
On macOS Sequoia, Tahoe and later:
- Open the
.dmg(Apple silicon for Macs with M1 or later, Intel for older ones) and drag Crow into Applications. - Open Crow from Applications. macOS says it could not verify Crow: click Done, not Move to Trash.
- Open System Settings → Privacy & Security and scroll down to Security. Next to "Crow" was blocked to protect your Mac, click Open Anyway, and confirm with your password or Touch ID.
- macOS asks once more: click Open. From now on Crow opens with a double-click.
- No Open Anyway button? Open Crow again, then go straight back to Privacy & Security.
- Control-click → Open no longer skips this check since macOS Sequoia: use Privacy & Security.
- If macOS says Crow "is damaged and can't be opened", open Terminal, run
xattr -dr com.apple.quarantine /Applications/Crow.app, and open Crow again.
Then macOS asks once per folder (Documents, Desktop, Downloads) before Crow may look into it: allow the folders you want in your second brain.
Windows "Windows protected your PC"
On Windows 10 and 11:
- Download
Crow_…_x64-setup.exe. If Microsoft Edge says it isn't commonly downloaded, click … next to it → Keep → Show more → Keep anyway. In Chrome, click Keep. - Run it. When Microsoft Defender SmartScreen says Windows protected your PC, click More info → Run anyway.
- Blocked with no Run anyway button? That is Smart App Control, which blocks every unsigned app on Windows 11 and has no exception for one app. You can turn it off in Windows Security → App & browser control → Smart App Control settings; on some Windows 11 versions it cannot be turned back on without reinstalling Windows, so decide with that in mind.
Linux AppImage, .deb, .rpm
- AppImage, any distribution:
chmod +x Crow_*.AppImage, then./Crow_*.AppImage(or right-click → Properties → Allow executing file as program). If it doesn't start and mentions FUSE, install it:sudo apt install libfuse2t64on Ubuntu 24.04 and later,sudo apt install libfuse2on 22.04. - Debian, Ubuntu:
sudo apt install ./Crow_*_amd64.deb - Fedora, openSUSE:
sudo dnf install ./Crow-*.x86_64.rpm(orsudo zypper installon openSUSE) - A blank window? Start Crow with
WEBKIT_DISABLE_DMABUF_RENDERER=1.
Start
- Pick the folders Crow should read, or single files.
- Press Connect: OpenRouter opens in your browser. Sign in (or make an account), press Authorize, and Crow gets a key of its own on your account, which you can give a credit limit or an expiry.
- Choose a model, check the cost, and start. You can chat with the wiki while it grows.
Already have an OpenRouter key? Crow also reads OPENROUTER_API_KEY from the environment, or from a .env file.
What it costs
Crow is free. You pay OpenRouter for the models it uses. Before filing starts, Crow estimates the cost from OpenRouter's current prices and the real size of each document you picked, and it stops on its own at your spending limit ($10 unless you change it). With the default model, a typical document costs a fraction of a cent; a library of thousands takes hours, so start with a folder or a recent date range and add more later.
Privacy
Your wiki and your chats are stored only on your computer, but what the models read goes to OpenRouter, which passes it on to each model's provider: to file a document, its text, its name, where it is on your computer and the notes it may join; to answer a question, the question, the notes the answer draws on and the chat's earlier messages. What they keep is up to OpenRouter's privacy policy and each provider's own. Your privacy settings on OpenRouter let you turn away providers that train on prompts.
Films
Crow on film
The films posted on LinkedIn and X since the launch.
The engine
LLM Wiki v2
Crow runs the CROW architecture of the open-source llm-wiki-v2 engine. Give it your documents: it files them into a tidy wiki of Markdown notes and answers your questions, citing the notes it used. It runs on your computer, with any AI provider: OpenRouter, OpenAI, Gemini, Ollama (free and local) or any OpenAI-compatible server.
It is a Python library, a command line and a web page with the brain, and it can be the long-term memory of Claude Code, Codex or any MCP client. Every step of the ingest and ask flows is a function you can replace.
crow-hq/llm-wiki-v2254 stars · 41 forks on GitHub
llmwiki2 on PyPIVersion 0.8.0 · Python 3.11 or later
Library guide
from llmw2 import Wiki
Memory for AI agents (MCP)recall, remember, read_note
How CROW decidesThresholds and fallbacks
BenchmarksCorpora and scripts, with --dry-run
ChangelogEvery release since 0.2.0
DiscussionsQuestions and ideas
Start in 2 steps
pip install llmwiki2 llmwiki2
Your browser opens: pick a provider, paste its key, press Test connection, then Save and start. No Python? uv tool install llmwiki2.
Just looking: the example wiki, no key needed
uvx llmwiki2 demo
Claude's memory (MCP)
pip install 'llmwiki2[mcp]'
# open Memory on the page, press Start, then:
claude mcp add --transport http --scope user llm-wiki http://127.0.0.1:8000/mcp
codex mcp add llm-wiki --url http://127.0.0.1:8000/mcp
As a Python library
from llmw2 import Wiki
wiki = Wiki.from_env(bundle="./wiki")
wiki.init()
wiki.ingest_file(open("minutes.pdf", "rb").read(), "minutes.pdf")
answer = wiki.ask("What did we decide about enterprise pricing?")
print(answer.text, answer.citations)
llmwiki2's web page: a PDF dropped on the page is merged into the note it updates; a typed question gets an answer citing that note; a second document dropped on the brain appears as a new note in its folder
Research
The paper behind it
CROW: Classifier-Routed Organization of LLM Wikis
LLM wikis have been proposed as an alternative to classical retrieval-augmented generation (RAG): a large language model (LLM) incrementally organizes knowledge into a directory of notes that can be inspected directly and grows with little manual upkeep, at the price of LLM calls whose cost grows with the knowledge base.
Many of these calls are spent on decisions rather than on writing: where a new source belongs, which notes it relates to, which notes answer a question. We introduce CROW (Classifier-Routed Organization of LLM Wikis), a decision layer for LLM wikis that delegates these routing and consolidation decisions to a general-purpose typed classifier, such as the recently released Jev, and reserves generative calls for the steps that require writing text.
In CROW, ingestion relies on a short sequence of classifier calls and invokes an LLM only to summarize the incoming source, to name new folders and, when an existing note must be updated, to rewrite it, while retrieval uses the classifier to select the relevant notes and leaves to the LLM only the final answer. CROW does not replace the compilation of entity, concept and synthesis pages that gives LLM wikis much of their value. We specify how each decision maps onto classifier primitives, position the design with respect to memory systems for LLM agents, model routing and candidate-then-verify pipelines, and discuss the effects on cost and latency that we measured in a 525 documents ingestion and retrieval benchmark.
News
Six weeks, no ads
From the paper on 22 September to the code going public on 9 October, the desktop app a week later and a release every few days since. The numbers below are the ones published on LinkedIn on 3 November 2026.
140,458impressions on the launch post on LinkedIn
250+stars on GitHub, 254 today
40+forks, 41 today
~3,000downloads from PyPI, 661 in the last week
~3,000views of the paper on Zenodo
Timeline
llmwiki2 0.8.0
Wikis of thousands of documents built from empty: top-level areas you decide, costs estimated and capped, steady speed as it grows, and nothing lost if a run stops halfway.
Crow 0.1.4 and llmwiki2 0.7.0: Connect AI apps
Claude, Claude Code, ChatGPT and Codex use the wiki as their memory, through MCP, with one button each.
llmwiki2 0.6.0 and Crow 0.1.3
Each document is filed with one choice among all the wiki's folders, which costs less, and a folder that grows is split into subfolders by subject.
Crow 0.1.0: the desktop app
For macOS, Windows and Linux, over llmwiki2 0.4.4, with an onboarding in five screens. The same day, Crow 0.1.1 opened a note's documents and deleted notes and folders, and Crow 0.1.2 moved to llmwiki2 0.5.0, which reads Word, PowerPoint, Excel and OpenDocument files itself.
pip install llmwiki2The first release on PyPI, 0.4.1.
llmwiki2 0.4.0
Move notes, describe folders, ingest large documents; CROW works out the classifier's confidence itself.
The code goes public: LLM-Wiki-V2
crow-hq/llm-wiki-v2 on GitHub, and a new Zenodo record of the paper's version 1.0 that links to the code.
The paper's version 1.0
On Zenodo, after the preprint.
The CROW preprint
CROW: Classifier-Routed Organization of LLM Wikis, on Zenodo.
Others writing about CROW
On LinkedIn and X
inWe just released CROW 0.8.0It keeps a wiki of thousands of documents in order, with areas you decide, costs estimated and capped, steady speed as it grows, and nothing lost if it stops halfway.
inBuilding a brain is the new hot topic, and we made it accessible to everyoneWe made it easy, with no scaling limit, and gave it to everyone for free.
inTurn your messy documents into knowledge your AI agents can useOrganize all the documents on your laptop into a clean wiki, then connect it to your favorite agent with one click: Claude, Claude Code, ChatGPT or Codex.
inYour desktop. Total chaos.Downloads. PDFs. Notes you'll never find again. One click. … Everything sorted. Everything linked. Your second brain, built for you.
inCROW 0.6.0 is out: smarter folders, lower costThe new version builds real folders and subfolders on its own, at a fraction of the cost: 0.86× the cost of CROW 0.5.0, 0.45 to 0.65× the cost of classic LLM filing.
inBuild your second brain in 7 clicks with CROWWe are releasing Crow-Desktop, an interface for the CROW architecture that lets you build your LLM wiki on your machine, with no scaling limits. Best part is, it detects all the documents by itself.
inpip install llmwiki2That's all. Now you can build your personal LLM wiki on your device, and it stays fast as it grows.
inCROW lets a small typed classifier take the filing decisionsCROW now works out that confidence itself, so the same answer takes the same decision on a hosted API, on your own server, or on a laptop with no network.
inOur LLM-Wiki has no scaling limitsWith the latest release of CROW, we filled the remaining gaps, so you can finally bring your LLM wiki to any production environment.
inLLM wiki is back: introducing LLM-Wiki-V2The great thing about an LLM wiki, compared to classical RAG, is that it reorganizes your documents into a structure that both agents and humans can read: a browsable tree of folders instead of an opaque vector index.
inWe brought the LLM wiki back from the deadA typed classifier (like Jev or Laya) makes the filing decisions, and the LLM only writes. It runs using Google Open Knowledge Format OKF.
inCROW: Routing LLM Wikis Without the LLM: Cheaper, Faster, BetterMe and Marco Sassarini propose CROW (Classifier-Routed Organization of LLM Wikis), a decision layer that delegates routing, retrieval and consolidation to a general-purpose typed classifier.
XCROW 0.6.0 on X
XCROW and Crow Desktop on X
Made by
Federico Cesarini and Marco Sassarini
The two authors of the CROW paper, of llm-wiki-v2 and of Crow.
Federico Cesarini
I build memory systems for agents, and humans.His bio on X
Marco Sassarini
AI engineer. Agents, RAG and LLM pipelines measured on real data, not demos. Open source • code that actually shipsHis bio on GitHub
Questions, ideas, bugs, or a commercial licence for llm-wiki-v2?
FAQ
Questions
What is a second brain?
A second brain is a system outside your head that keeps what you read and learn, so you can find it and connect it later. Crow builds one from the documents you already have: you don't write, tag or file anything yourself.
What is an LLM wiki?
An LLM wiki is a set of Markdown pages that a language model writes and keeps up to date from your sources, an idea Andrej Karpathy described in 2026. Instead of searching your documents again at every question, the model reads each source once and files what it says. Crow is a desktop app that builds one from your files and keeps it organized as it grows to thousands of documents.
How do I chat with thousands of PDFs and documents on my computer?
Install Crow, tick the folders that hold them, and start: Crow files them into the wiki, oldest first, and you can ask questions while it works. Each answer names the notes it used, so you can open them and check.
Can I use it with Obsidian?
Yes. The wiki is a folder of Markdown files with relative links and an index.md in every folder, so Obsidian opens it as a vault, graph view included. Settings → Library → Show in Finder/Explorer shows where it is. Crow doesn't need Obsidian.
Does Crow change, move or upload my files?
It never changes or moves them: Crow reads the folders you choose and keeps a copy of each document's text inside the wiki. Deleting a note, or resetting Crow, never touches your documents. To file a document, its text goes to the model you chose through OpenRouter (see Privacy, under Download).
Which AI models can I use? Can I run it offline or with Ollama?
Gemini 3.8 Flash (the default), Claude Sonnet 5.5, DeepSeek V4.1 Flash, or any other model on OpenRouter, under Advanced settings; a small classifier model, Jev, decides where each document goes. Crow works through OpenRouter, so it needs an internet connection; local models are not supported in the app yet. The llm-wiki-v2 engine runs with Ollama from the command line.
Can my AI agents use the wiki as memory?
Yes: Connect AI apps, in Crow's sidebar, connects Claude, Claude Code, ChatGPT and Codex in one click each (see AI apps). Any other MCP client (Cursor, VS Code…) can run the same local (stdio) server: Crow's own executable with the argument mcp. The wiki is also plain Markdown with an index in every folder, so an agent that reads files can use it directly.
How is Crow different from NotebookLM?
NotebookLM works on the sources you upload to it, notebook by notebook, in Google's cloud. Crow finds the documents on your computer by itself and writes one wiki of Markdown files that stays on your computer, which you can read, edit and keep, with the models you choose through OpenRouter.
What is Jev?
A typed classifier by TypeSafe AI. Given a text, it answers typed questions (pick one of these options; is this statement true?) with probabilities instead of prose. CROW asks it every filing and retrieval decision and keeps the LLM for writing. CROW does not depend on it: any classifier that returns those probabilities can take its place. The authors have no ties to TypeSafe AI.
Why does my Mac or PC say Crow can't be verified?
Because Crow's installers are not signed with a paid Apple or Microsoft developer certificate yet. The warning is expected; Install Crow shows how to open it.
Why AGPL?
The AGPL keeps the code and every change to it open, also when it runs as a service. It is free to use, study, change and share, including for your own business; if you offer a modified version to others, over a network too, you share its source under the same licence. For llm-wiki-v2, commercial licences without that obligation are available from the authors: open an issue to ask.
Something wrong?
- Logs: Settings → About → Open logs.
- The wiki itself: Settings → Library → Show in Finder/Explorer, or right-click the brain's empty space.
- Start over: Settings → Library → Reset Crow… erases the wiki, the chats and the settings. Your documents stay where they are.
- Bugs and ideas: issues.