Installation and quick start
This takes about 5 minutes if your API key is already handy, 5 to 10 if you let Vault Operator download a local embedding model.
Before you start
- Obsidian 1.13 or later on macOS, Windows, or Linux. Vault Operator is desktop-only. Earlier versions are blocked by
manifest.json. - An internet connection if you plan to use a cloud model. If you want everything offline, install Ollama or LM Studio first.
- One AI provider API key. Free options exist (Google Gemini, Ollama). API keys look like
sk-ant-...for Anthropic,sk-...for OpenAI, or a long random string for Google. Get one from the provider's dashboard before you start. - About 100 MB of disk space for the plugin, the embedding model, and the knowledge database. Large vaults add more.
If something goes wrong, the Troubleshooting page covers the common failures (connection errors, semantic search not working, agent stuck in a loop).
Install the plugin
Vault Operator is available in the official Obsidian Community Plugins directory.
- Open Settings > Community plugins > Browse
- Search for Vault Operator (plugin id
vault-operator) - Click Install, then Enable
The Vault Operator icon appears in the left sidebar.
Shortcuts
- Direct deep link (opens Obsidian): obsidian://show-plugin?id=vault-operator
- Community page: community.obsidian.md/plugins/vault-operator
Run the first-run wizard
A seven-step setup wizard opens automatically the first three times you enable the plugin. The steps cover welcome, provider, model selection, embeddings, default folder, agent folder, and finish. It walks you through every step on this page.
To rerun the wizard later, open Settings > Vault Operator > Advanced > Interface > Setup > Restart setup.
If you prefer to set things up manually, read on.
Add your first provider
Vault Operator needs an AI provider to work. Open Settings > Vault Operator > Providers > Providers and click + Add provider. The Add provider modal opens.
Free option (no credit card)
- Go to Google AI Studio
- Sign in and click Create API key
- In the Add provider modal, pick Google Gemini as provider type and paste the key
- Click Refresh next to the Models field. Vault Operator pulls Gemini's model list and sorts it into Budget, Main, and Frontier tiers.
Google Gemini has a free tier with reasonable rate limits, which is enough to try everything out before paying anyone.
Best quality
| Provider | Default main | Default frontier | Notes |
|---|---|---|---|
| Anthropic | Claude Sonnet 4.5 | Claude Opus 4.6/4.7 | Best tool use in testing |
| OpenAI | GPT-5.1 / GPT-4.1 | GPT-5, GPT-5-pro | Fast, good at structured output |
| Gemini 2.5 Flash | Gemini 2.5 Pro | Free tier, large context window |
Local and private
If you want no data leaving your machine, run a model locally:
- Install Ollama, then run
ollama pull llama3.2 - Or download LM Studio, install a model, and start the server
For local providers the Base URL field pre-fills with the default port. Click Refresh next to the Models field and you are done.
Multiple providers
Vault Operator supports 12 providers. You can configure several and switch between them through the Active provider radio, or pin a specific model in the chat header for a single task.
Your first chat
- Click the Vault Operator icon in the left sidebar
- Type a message and press Enter
- Watch the agent work. It shows every tool call in real time
Try these prompts
- "What notes do I have about [any topic]?"
- "Summarize the note I'm currently viewing"
- "Create a new note with a summary of my last 3 daily notes"
- "Find all notes tagged with #project and create a canvas showing their connections"
Inline AI chat (v3.0.0)
Select text in any note and press Cmd+Shift+I (Ctrl+Shift+I on Windows and Linux) to open the inline AI chat panel directly over the selection. See Inline chat for details.
What happens behind the scenes
When you send a message, Vault Operator reads it and decides which tools to use. It then calls those tools (read files, search, write) while you see each call in the activity block. Before any write operation it asks for your approval, unless you've enabled auto-approve for that category. Then it returns a response.
Every write operation creates a checkpoint, so you can undo any change with one click.
By default the chat loop runs on the Main tier of your active provider. When the agent hits a hard synthesis step (cross-document reasoning, ambiguous instructions) it can escalate one step to the Frontier tier via consult_flagship, capped at three calls per task. The cost log in the developer console tags each turn with mode=auto, mode=advisor, mode=override, or mode=subagent. See Advisor pattern for details.
Next: Your first conversation
Continue with Your first conversation to learn context, approvals, and the activity block.
Then walk the knowledge tutorials in order:
For deeper references, see Choosing a model and Safety and control.
If something does not work as described above, check the Troubleshooting page first.