Vault

Chat with your notes on your own computer. Every answer cites the note it came from, or says it doesn't know.

Vault (Local Knowledge Vault) is a desktop app for macOS, Windows and Linux. Your notes stay in a local SQLite database. Answers come from a local model by default (Ollama or LM Studio, detected automatically), and you can add a cloud provider or a shareable plugin in a few clicks.

_{An answer with its citation badge. (This still used a cloud provider; new installs use a local model first.)}

Download

Installers will be published on the Releases page: .dmg/.zip for macOS, an NSIS .exe for Windows, and .AppImage/.deb/.rpm/.snap for Linux. Until the first release is up, build from source.

Linux: on Ubuntu/Debian run sudo apt install ./local-knowledge-vault_*_amd64.deb; on Fedora run sudo dnf install ./local-knowledge-vault-*.x86_64.rpm. The snap is confined, so Media chat can't reach a yt-dlp installed outside it; use the .deb, .rpm or AppImage if you need YouTube import.

macOS: the app is not signed yet. The first time you open it, macOS will block it:

• Right-click (or Control-click) Vault.app → Open → Open, or
• Try to open it once, then go to System Settings → Privacy & Security and click Open Anyway (on recent macOS versions this is the only route).

Quick start (local first)

1. Install Ollama or LM Studio.
2. Pull the recommended model:

   ollama pull qwen3:8b

   (about 5 GB; llama3.1:8b also works. In LM Studio, search for "Qwen3 8B" and load it.)
3. Open Vault. The header chip should read Local · Ollama · qwen3:8b. If no model is running, Vault shows a setup card with the command above and a Re-check button.
4. Add notes: New note, or turn on Advanced and paste a link into Add from URL. First launch also seeds a few sample notes so you can try it straight away.
5. Open Ask and ask a question. Click a citation badge to open the source note.

Prefer a cloud model? Click the AI chip → Add provider, pick a preset, paste your key, press Test connection, then Save.

Why Vault

• Answers you can check. Answers cite the notes they came from as clickable itm_ badges. Citations to notes that weren't retrieved are dropped before you see them, and when your notes don't cover a question, Vault says so.
• Local by default. Notes live in SQLite on your machine. With Ollama or LM Studio, nothing leaves your computer. Search works with no model and no network at all.
• Any model you like. Local models first, or OpenAI, Anthropic, Gemini, OpenRouter, Mistral, DeepSeek, Together, Groq, xAI, or any OpenAI-compatible URL.
• Plugins you can share safely. A plugin is a plugin.json folder or zip. It adds presets, voices and prompt packs, and it cannot run code.
• Low ceremony. Ask is the home screen. Simple mode hides the knobs; Advanced shows them.

Features

Ask with citations

Vault searches your notes, sends the matching passages to the model with fixed grounding rules, and shows each cited note as a badge. Clicking a badge opens the note in a side peek, so your chat stays where it is. If search finds nothing, you get "I couldn't find that in your notes" without a model call. Chats are saved as sessions, and Export citation pack writes the thread, the cited notes and a manifest to a folder or zip that you can hand to another LLM or a colleague.

Local model setup that explains itself

With no model running, the first-run card recommends qwen3:8b and offers Re-check or Use a cloud model instead. If you are running a very small model (under ~3B parameters), Vault suggests a bigger one, because small models follow citation rules less reliably.

Add your own provider

Pick a preset or a custom OpenAI-compatible URL, fetch the live model list, and test the connection before saving. Keys are write-only in the UI and stored in owner-only files.

Declarative plugins

Plugins → Manage plugins installs a folder or .zip, shows what each pack contributes, and explains why a broken pack was rejected.

More

• Media chat. Paste a YouTube URL (captions fetched with yt-dlp, which must be installed; pipx install "yt-dlp[default,curl-cffi]" includes the browser impersonation that avoids YouTube 429 errors) or open a local video/audio file with its .srt/.vtt captions. Captions become timed transcript notes you can ask about. With local media, clicking a citation seeks the player; YouTube embeds show the timestamp instead. Player and chat can go fullscreen together.
• Media voices. Reusable voices (Desk cohost, Curious student, Skeptical investor, or your own) that change the speaking style but keep the same grounding rules. One install works for every video.
• Personalities and profiles for Ask, plus note groups (Projects / Areas / Resources / Archives), kinds, status and project filters.
• Add from URL. Fetches a public web page, extracts the text and auto-tags it as a note (Advanced mode).
• MCP client. Plugins → MCP connections → Connect Notion signs in to Notion's hosted MCP server (https://mcp.notion.com/mcp) with OAuth and lists its tools. You can also add other Streamable HTTP MCP servers. Using MCP tools inside Ask is not wired up yet.
• Obsidian bridge. While Vault runs, it serves a small HTTP API on 127.0.0.1:8765 (/health, /v1/projects, /v1/ask). A scaffold Obsidian plugin in bridges/obsidian-vault/ adds an Ask Vault… command that inserts the cited answer into your current note. See docs/bridges.md.
• Simple and Advanced modes, light and dark themes.

Retrieval is local keyword search (SQLite FTS5 with BM25 ranking). There are no embeddings, so short, focused notes and questions that reuse their words work best.

Bring any model

Preset	Default model	Key
Ollama (local)	auto-picks an installed model	none
LM Studio (local)	the loaded model	none
OpenAI	gpt-4o-mini	required
Anthropic	claude-haiku-4-5	required
Google Gemini	gemini-3.8-flash	required
OpenRouter	openrouter/auto (openrouter/free for free models)	required
Mistral	mistral-small-latest	required
DeepSeek	deepseek-flash	required
Together	meta-llama/Llama-3.3-70B-Instruct-Turbo	required
Groq	openai/gpt-oss-20b	required
xAI	grok-4.3	required
Custom	any OpenAI-compatible server (vLLM, llama.cpp server, LocalAI, Jan, gateways)	optional

In Auto (local first) mode Vault tries Ollama, then LM Studio, then other local servers, then enabled cloud providers that have a key. A disabled cloud provider is never called. Model defaults change often, so use Fetch models to pick from the provider's live list. Details, key storage and environment variables: docs/providers.md.

Plugins

A plugin is a folder with a plugin.json (plus optional Markdown and image files). It can add provider presets, voices, quick-ask prompt packs and MCP server presets. It cannot run code, and it cannot ship API keys or custom headers.

{
  "schemaVersion": 1,
  "id": "study-buddy",
  "name": "Study buddy",
  "version": "1.0.0",
  "contributes": {
    "personas": [
      { "name": "Study buddy", "prompt": "Explain simply, then ask one quiz question from my notes." }
    ],
    "promptPacks": [
      { "name": "Revision", "prompts": ["Quiz me: ask 3 questions I should be able to answer from these notes."] }
    ]
  }
}

Install it from Plugins → Manage plugins → Install folder… (or Install .zip…). Full reference: docs/plugins-authoring.md. Examples: examples/plugins/. Built-in panels are documented in PLUGINS.md.

How it compares

	Vault	Obsidian + AI community plugins	Notion AI	AnythingLLM
Where your notes live	SQLite on your computer	Markdown files on your device	Notion's cloud workspace	On your computer (desktop) or your own server
AI	Built in: grounded Ask over your notes	Through community plugins; behavior varies by plugin	Built in (Business/Enterprise plans; limited trial on Free/Plus)	Built in: chat with documents, agents
Models	Local first (Ollama, LM Studio), or add a cloud provider	Depends on the plugin	Models Notion offers, via its model picker	Local on-device models, with model selection
Best at	Cited answers from personal notes and video captions, one user	Flexible linked notes, thousands of plugins, sync and publish	Team docs, databases and collaboration	A broad local AI workspace, including multi-user self-hosting

Where Vault is weaker today: no semantic (embedding) search, single user, no mobile app, no sync, and a much smaller plugin ecosystem. Obsidian and Vault can work together through the bridge.

Privacy

Your data lives in Vault's user-data folder:

OS	Folder
macOS	~/Library/Application Support/local-knowledge-vault
Windows	%APPDATA%\local-knowledge-vault
Linux	~/.config/local-knowledge-vault

It holds the notes database (lkv.sqlite), provider settings (lkv-providers.json), API keys (lkv-keys/, owner-only files) and plugins (plugins/). Set LKV_USER_DATA_DIR to use another folder.

What leaves your machine:

• With a local model: nothing. Health checks only ping local servers.
• With a cloud provider you enabled: your question, the matching note passages and recent chat turns go to that provider only. Cloud providers are not contacted until you ask something.
• Add from URL fetches the page you give it. Vault asks the active model to auto-tag it, so if that is a cloud provider it receives an excerpt of the page.
• Media chat contacts YouTube (through yt-dlp and the embedded player) only for YouTube media.
• Notion is contacted only if you connect it under MCP connections.
• The Obsidian bridge listens on 127.0.0.1 only and requires a per-install bearer token (shown in AI providers → Vault Bridge). Do not expose or port-forward that port.

Build from source

Requires Node.js and npm. better-sqlite3 is a native module; on Linux you may need python3, make and g++.

git clone https://github.com/tosin2013/local-knowledge-vault.git
cd local-knowledge-vault
npm install          # also rebuilds better-sqlite3 for Electron
npm run dev          # Electron + Vite with hot reload

Checks (no GUI needed):

npm run typecheck
npm run test:mvp            # database, search, citation validation
npm run test:providers      # provider registry, adapters, plugin loader (offline)
npm run test:media          # caption parsing and chunking
npm run test:citation-pack  # citation pack export

Installers:

npm run dist:mac     # dmg + zip (on macOS)
npm run dist:win     # NSIS installer (on Windows)
npm run dist:linux   # AppImage + deb + rpm + snap (on Linux)

More (layout, IPC API, environment variables, all test scripts): docs/development.md. Releases: docs/release.md.

Roadmap

• Vault as an MCP server, so AI agents can use your notes as cited memory.
• Signed and notarized macOS builds.
• A starter sample vault for a better first run.
• Code plugins later, behind a permission model. Plugins stay declarative for now.

Contributing

Issues and pull requests are welcome. Please include your OS, the model or provider you used, and steps to reproduce. Before opening a PR, run npm run typecheck and the relevant npm run test:* scripts. For plugin ideas, start with a plugin.json pack; see docs/plugins-authoring.md.

License

Vault is licensed under the Apache License 2.0. See NOTICE for attributions.

Built by Tosin Akinosho (@tosin2013).