Self-hosted LLM Vault Agent

Chat with your own LLM server from an Obsidian sidebar and let the model read, search and write your notes. Works with Ollama, vLLM, LM Studio, llama.cpp server, LiteLLM, or any server that speaks the OpenAI-compatible /v1/chat/completions API — on your machine, in your home network, or at your university/company.

Prefer a cloud model? The sibling plugin Euridian connects to Infomaniak Euria (repo). Both plugins share the same core.

⚠️ Vault agent — please read before enabling

The plugin can give the AI model function-calling tools that act on your vault: list notes, read notes, search, create notes, append to notes, and overwrite a note's entire content. This means the model can modify your files.

• No delete tool exists at all — not locked, simply not implemented.
• Writes require confirmation by default (setting Confirm before writing): every create/append/overwrite shows a preview modal (with a word diff for overwrites) before anything touches disk. You can turn this off, but it's not recommended.
• Overwriting a note replaces its entire content. Anything the model doesn't repeat back is lost. The modal highlights this in red when content would shrink.
• The agent can be disabled entirely (Vault agent setting) if you only want plain chat.
• Note content you ask about is sent to the server you configured. With a server on your own machine nothing leaves it. See Privacy below.

✨ Features

• 💬 Chat sidebar with multiple tabs, persistent history, streaming responses; each chat keeps its own backend and model
• 🤖 Vault agent (function calling) — read, search, create, append and edit notes on request, with confirmation + diff preview before any write
• 📄 Documents — the agent can also read PDF, DOCX and PPTX files in your vault
• 🌐 Optional web search — DuckDuckGo (no account) or Brave Search API, off by default
• 📎 Attachments — drag & drop files/images, attach vault notes or the current editor selection
• ⚡ Slash commands — reusable prompt templates (/summarize, /translate, …) with placeholders {{input}}, {{selection}}, {{note}}, {{title}}
• 🔗 @mention — reference any vault note from the chat input
• ✏️ Inline edit — select text, edit it with AI, review a word diff, accept or reject
• 📝 Custom instructions file — a short markdown file in your vault with conventions for the agent (see below)

Language: the interface is in English and follows Obsidian's language setting; a German translation is included. Built-in prompt templates and prompts to the model are localized the same way.

🤖 Using the vault agent

With Vault agent enabled (default) and a model that supports function calling, just ask naturally:

"Summarize the note 'Project X'" "Search my vault for notes about network routing" "Append today's meeting notes to my daily note" "What does the PDF in my attachments folder say about warranty terms?"

Each tool call appears as a chip in the chat. Extracted document text is capped at 40k characters per file; scanned PDFs without a text layer return no text. Not every model supports function calling reliably; very small models (≲1B parameters) tend to hallucinate results instead of calling tools.

🌐 Web search (optional)

Adds a search_web tool. It is off by default, independent of the vault agent, and always runs locally through your own internet connection — so it also works when your model runs on a server without internet access. Enable it in Settings → Enable web search and pick a provider:

• DuckDuckGo (default, no account or key) — reads DuckDuckGo's HTML results page. This is unofficial, so it can break or be blocked with a bot check after many requests.
• Brave Search API — more stable, needs a free API key from brave.com/search/api (free tier available).

📝 Custom instructions file

Point the plugin at a short markdown file (Settings → instructions file, default Agent.md in the vault root) describing your vault's structure and conventions. Keep it short and specific: large general-purpose instruction files written for another assistant can overwhelm smaller models so that they stop calling tools.

🚀 Quick start

1. Run a server. For Ollama:

   brew install ollama
   ollama pull qwen3      # or any model with tool-calling support
   ollama serve
   

2. Enable the plugin under Settings → Community plugins, open its settings, and enter the Server URL (default http://localhost:11434; for other servers e.g. https://llm.example.org, without a path). Add an API key only if your server requires one.
3. Click Scan models to list the server's models (or type a model name), then Test to verify the connection.
4. Open the chat: Command palette → Vault Agent: open chat, or use the ribbon icon.

If the server sits behind a VPN or internal network, connect before testing.

🔐 Privacy & network use

The plugin is desktop-only and sends requests only to the servers you configure:

No telemetry, no other network calls. The optional API key is stored in plain text in Obsidian's data.json (standard Obsidian behavior). Server operators can see everything you send — for a server run by someone else (an institution, a company), check their data-handling rules before sending sensitive content.

🔄 Coming from Euridian?

Until version 0.3, Euridian also offered Ollama and custom servers. To take your settings and chats along, close Obsidian and copy .obsidian/plugins/euridian/data.json to .obsidian/plugins/self-hosted-llm-vault-agent/data.json; on first start the plugin converts the Ollama/custom settings into the server settings.

⚠️ Known limitations

• Reasoning models can "think" for minutes; the Thinking / reasoning toggle is off by default.
• Some servers cap the usable context well below the model's maximum; if requests fail with context-length errors, narrow the request or ask the operator.
• Servers or proxies that drop connections silently: the plugin aborts a stream after 90 s of complete silence and shows an error.

ℹ️ Requirements and disclosures

• No account or payment required by the plugin. You need access to an OpenAI-compatible server (your own, or one provided to you). Web search with Brave needs a free Brave Search API key; DuckDuckGo needs nothing.
• Network use. See the table under Privacy & network use. Nothing is sent anywhere until you send a message, load the model list, or run a web search.
• No telemetry, no ads, no self-updates.
• Not affiliated. The plugin is independent of Ollama, vLLM, LM Studio, DuckDuckGo and Brave; the names only describe compatible software.

📦 Third-party software

The plugin bundles these libraries (their licenses apply):

The bundled main.js is minified but not obfuscated; the full source is in this repository. pdf.js contains a one-time feature check (new Function("")) that the plugin does not use for anything else; the plugin itself runs no dynamically generated code.

🔧 Development

git clone https://github.com/jlueschow/self-hosted-llm-vault-agent
cd self-hosted-llm-vault-agent
npm install
export EURIDIAN_PLUGIN_DIR="/path/to/YourVault/.obsidian/plugins/self-hosted-llm-vault-agent"
npm run dev      # watch, rebuild and deploy on every change
npm run build         # production build (dist/main.js)
npm run build:deploy  # production build + copy into your vault

src/core is a copy of the shared core from the Euridian repo — do not edit it here. Update it with node scripts/sync-core.mjs ../obsidian-euridian-plugin; CI checks for drift. src/variant holds what is specific to this plugin.

📄 License

MIT — see LICENSE.