gknowledge

gknowledge

English · 简体中文 · Italiano · Français · Português (BR) · Español

Karpathy's LLM Wiki as a native GNOME app.

Drop in your documents. Get a linked wiki that an LLM keeps up to date.

A native GNOME app. Offline by default, plain Markdown on disk, Obsidian-compatible.

Bring your own brain: a local model, any OpenAI-compatible endpoint, or your own ChatGPT plan.

Real output of the offline engine on six AI papers ( data/demo/papers ), no API key. Regenerate with make demo-svg .

Install

One line, on Ubuntu 24.04+ and Debian 13+. It adds the signed gknowledge APT repository, so updates arrive with apt upgrade (other distributions get a per-user install):

curl -fsSL https://github.com/openapi/gknowledge/releases/latest/download/install.sh | bash

Or add the APT repository yourself. apt also pulls GTK 4, Libadwaita and PyGObject:

sudo curl -fsSLo /usr/share/keyrings/gknowledge.gpg https://openapi.github.io/gknowledge/gknowledge.gpg

sudo curl -fsSLo /etc/apt/sources.list.d/gknowledge.sources https://openapi.github.io/gknowledge/gknowledge.sources sudo apt update && sudo apt install gknowledge

On any distribution, with Flatpak (the GNOME runtime comes from Flathub):

flatpak install --user https://openapi.github.io/gknowledge/io.github.openapi.Gknowledge.flatpakref

Both repositories are described at openapi.github.io/gknowledge.

Then open gknowledge from the app grid. To remove it: sudo apt remove gknowledge (or flatpak uninstall io.github.openapi.Gknowledge).

Other distributions, from source, and release builds

• Fedora, Arch and others: install PyGObject, GTK 4 and Libadwaita with your package manager, then run the one-line installer above. It installs into ~/.local without root.
• From source: git clone https://github.com/openapi/gknowledge && cd gknowledge && make install.
• Single files: every version's .deb and Flatpak bundle is on the Releases page, e.g. sudo apt install ./gknowledge_all.deb after downloading gknowledge_all.deb.

Why gknowledge

If you have used LLM Wiki, Tolaria, OpenKnowledge, Reor or Khoj, you already know the loop: ingest sources, let the model build and maintain a wiki, then ask questions against it. gknowledge does the same thing as a real GNOME application.

• Native, not a web view. GTK 4 and Libadwaita, built to follow the GNOME HIG. No Electron, no Tauri, no browser engine.
• Works with no account. A built-in heuristic engine ingests and answers offline. Add a model when you want better synthesis.
• Your data stays yours. Markdown with YAML frontmatter plus a readable SQLite index. Open the folder in Obsidian or any editor.
• Cited, reviewable, incremental. Every generated page records its sources, answers cite pages, and a review queue keeps you in control.

The workflow you already know

See ROADMAP.md for dates and TODO.md for the backlog.

Karpathy's LLM Wiki, native on GNOME

In April 2026 Andrej Karpathy described the LLM Wiki pattern: instead of re-reading raw documents on every question, an LLM builds and maintains a persistent wiki that compounds over time. gknowledge implements it as a desktop app.

What we kept

• Three layers: immutable sources/, the generated knowledge/ wiki, and settings that tell the engine how to work.
• The operations: ingest (read a source, update the pages it touches) and query (answer from the wiki, with citations). Lint is next.
• index.md as the catalogue and log.md as the append-only history.
• Plain Markdown with [[wikilinks]] and YAML frontmatter, so the folder opens in Obsidian.
• The human curates, the model maintains.

What we changed and added

• A native app instead of an idea file. GTK 4 and Libadwaita, with sources, wiki, assistant and review in one window.
• Works with no model at all. An offline engine ingests and answers; a model makes it better.
• Bring your own brain. Your ChatGPT plan through *Sign in with ChatGPT*, Ollama, or any OpenAI-compatible endpoint.
• Two-step ingest with provenance. Analysis first, then pages; every page records its sources.
• No empty pages, no duplicates. Pages are created only with real content, names are merged across sources ("Kasparov" and "Garry Kasparov"), minor names stay as mentions until a second source cites them.
• A review queue for what the model is unsure about.
• An MCP server, so Claude Code, Codex or Cursor can search, read and grow the same wiki.

Use your wiki from Claude Code, Codex or Cursor

gknowledge ships an MCP server. Point any agent at a project folder:

claude mcp add gknowledge -- gknowledge-mcp ~/Knowledge/my-project     # Claude Code

codex mcp add gknowledge -- gknowledge-mcp ~/Knowledge/my-project # Codex CLI

For Cursor or other clients, add this to the MCP configuration:

{ "mcpServers": { "gknowledge": { "command": "gknowledge-mcp", "args": ["/home/you/Knowledge/my-project"] } } }

Try it from source

sudo apt install python3-gi gir1.2-gtk-4.0 gir1.2-adw-1   # Ubuntu / Debian

git clone https://github.com/openapi/gknowledge && cd gknowledge make start # opens a demo project with six AI papers already ingested

make start also restarts the app whenever you edit the source, which makes it the fastest way to hack on gknowledge. Install it for daily use with make install.

Bring your own brain

Nothing is sent anywhere unless you pick a provider that is not offline.

How it compares

Details and history of each project are in DISTRIBUTING.md.

gknowledge is an alternative to…

gknowledge sits at the intersection of *LLM-powered knowledge bases*, *local-first note tools*, and *document RAG systems*. If you have used any of the tools below, you already understand the shape of the problem — gknowledge addresses it as a native GNOME/GTK application that is local-first, provider-agnostic, and usable fully offline (no AP