> A powerful AI assistant plugin that brings Claude, Gemini, OpenAI, Mistral, and local models directly into your Obsidian vault with intelligent note management capabilities.
> **New!** Restore file changes - Every note the AI creates, edits, or deletes (including binary files like PDFs and images) can be reviewed and restored to its previous or original version straight from the diff viewer.
- **Interactive Diff Viewer** - Review and approve AI-proposed changes before they're applied with side-by-side diff view, or let the AI make changes without asking"
- **Files Support (PDF, Documents, Images, etc)** - Search through PDF and Office documents (DOCX, PPTX, XLSX) and allow AI to read binary files directly (PDFs, Office/ODF documents, images, etc)
- **Quick Actions** - Lightweight, single-shot AI operations on selected text, available via right-click context menu and a dedicated editor toolbar button — both toggleable in settings and fully supported on mobile
When the AI proposes changes to your files in Agent Mode, an interactive diff viewer appears that lets you:
- **Review Changes**: See a side-by-side comparison of the current file and proposed changes
- **Approve or Reject**: Accept the changes to apply them, or reject to cancel
- **Provide Feedback**: Suggest modifications before accepting by typing your feedback and clicking the suggestion button
- **Mobile Support**: Touch-friendly Accept/Reject buttons on mobile devices
The diff viewer ensures you're always in control of what changes are made to your vault, providing transparency and safety when working with AI-generated edits.
You can toggle the agents ability to make changes without asking when in edit and planning mode. While enabled, proposed changes are applied immediately without the diff viewer prompt. Turn it off anytime to go back to reviewing every change.
Planning Mode introduces a three-agent workflow that separates task planning, orchestration, and execution. When enabled, specialized agents collaborate to analyze your vault, create a detailed strategy, and execute changes with intelligent oversight between each step.
6.**Plan Approval**: Review the plan and **Approve** it, **Reject** it, or suggest a change before execution begins (see [Plan Approval](#plan-approval) below)
7.**Execution Phase**: For each step, an execution agent performs the task while an orchestration agent monitors progress and decides whether to continue, adapt, or replan
8.**Completion**: All steps are marked complete when finished
- **Planning Agent**: Explores your vault (read-only) and creates the execution strategy. Can search files, read content, and ask clarifying questions, but cannot modify anything.
- **Execution Agent**: Performs individual steps from the plan. Has full access to vault operations (when in Agent Mode) and completes each task independently.
- **Orchestration Agent**: Monitors progress between steps. After each step completes, it evaluates the result and decides whether to continue to the next step, request a replan, or abort the workflow. It can also pass context to subsequent steps based on what was learned.
Before execution begins, the proposed plan opens in a dedicated view for your review:
- **Approve** - Accept the plan as-is and move to the execution phase
- **Reject** - Cancel the plan outright
- **Suggest a change** - Type feedback describing what you'd like changed, and the planning agent replans with your feedback in mind rather than starting over from scratch
If issues arise during execution, the orchestration agent can request a replan. This returns to the planning phase with full context about what's already been completed and what went wrong, allowing the plan to intelligently adapt to new information discovered during execution.
Quick Actions are one-click AI edits you run on the note you're currently editing, without opening the chat panel. Open the editor menu (right-click, or the command palette) and pick an action. Some actions work on your current selection if you have text selected, otherwise they apply to the whole note.
- **Proofread** — Corrects spelling, grammar, punctuation, and typos without rewriting for style or changing your voice. Works on the selection if you have one, otherwise the whole note.
- **Beautify** — Improves clarity, flow, and readability and adds Markdown formatting (headings, bold, lists, blockquotes) where it helps. Works on the selection if you have one, otherwise the whole note.
- **Apply template** — Restructures the note to match a template you choose, fitting your content to the template's headings and structure while preserving the information. If the file you pick doesn't look like a template, no changes are made.
- **Apply links** — Scans the note for mentions of pages that already exist in your vault and wraps them in wikilinks. It only links existing pages and never invents new ones. Works on the selection if you have one, otherwise the whole note.
- **Apply tags** — Chooses tags for the note from the tags that already exist in your vault and merges them into the frontmatter. It won't create new tags — only ones already in use elsewhere.
- **Suggest tags** — Like Apply tags, but free to suggest new tags as well as reuse existing ones. Suggested tags are merged into the note's frontmatter.
- **Generate frontmatter** — Infers YAML frontmatter for the note (aliases, tags, title, summary, created) from its content and merges it into any existing frontmatter.
The AI can retain information across conversation sessions — your vault conventions, tagging patterns, personal preferences, or any established workflows you've described.
**How It Works**
Memories are stored in a plain markdown file at `Vaultkeeper AI/memories.md`. The AI reads this at the start of each conversation and can update it during a session. You can view and edit this file directly at any time.
**Settings**
- **Enable Memories**: Toggle memory on or off entirely
- **Read-Only Memories**: The AI can recall past memories but cannot add or update them — useful if you want to manage the file yourself
### Web Search
Toggle web search on a per-message basis using the globe button in the chat input toolbar. When active, the AI can perform live web searches to answer questions requiring current or external information.
Web search is supported on all cloud providers (Claude, Gemini, OpenAI, and Mistral) but not on Local models. Mistral uses its native Agents API for search; Claude, Gemini, and OpenAI use their respective built-in search tools.
You can also disable web search access entirely from Settings if you prefer the AI to stay within your vault.
### Web Viewer
When enabled, the AI can fetch and read the content of web pages — useful for summarising articles, pulling in documentation, or referencing any URL you share.
Web viewer access can be toggled on or off in Settings.
Run Vaultkeeper AI entirely against a self-hosted, OpenAI-compatible server instead of a cloud provider — no API key required, and nothing leaves your machine.
**Setup**
1. Select **Local** from the provider dropdown in Settings (under the "Self Hosted" group)
2.**Local URL**: Enter your server's chat completions endpoint, e.g. `http://localhost:1234/v1/chat/completions` (LM Studio's default)
3.**API Key**: Optional — only needed if your local server requires authentication
4.**Model**: Type the name of the model to use, exactly as your server expects it
5. Optionally set separate **Planning Model** and **Quick Action Model** names if you want lighter/faster models handling those roles
This works with LM Studio, Ollama, vLLM, and other servers that expose an OpenAI-compatible chat completions API.
- PDF and Office/ODF document attachments (PDFs are automatically rasterized to images, since local models can't read raw PDF bytes) and text-based file attachments
**Limitations**
- No web search (cloud-only feature)
- Many local models need their chat template adjusted to reliably handle multi-turn tool-calling; check your server's documentation (e.g. LM Studio's prompt template settings) if tool use behaves unexpectedly
-`.obsidian/workspace.json` - Exclude specific configuration files
- **Common use cases**:
- Personal journals or diary entries
- Financial information
- Work-related confidential notes
- API keys or credentials stored in notes
- Draft content you don't want analyzed
- **Privacy guarantee**: Excluded files are completely inaccessible to the AI - they won't appear in searches, can't be read, and can't be modified even in agent mode
I'm currently **not accepting contributions** (pull requests are disabled), but bug reports and suggestions via [GitHub Issues](https://github.com/andy-stack/vaultkeeper-ai/issues) are always welcome. See [CONTRIBUTING.md](CONTRIBUTING.md) for details.
- **Fully Local Option**: Use the [Local](#local-models) provider with a self-hosted server to keep all inference on your own machine, with no cloud provider involved at all
- **File Exclusions**: Protect sensitive information by excluding individual files or entire directories from AI access using glob patterns - excluded files are completely inaccessible in both read-only and agent modes
- **Local Storage**: All conversations and settings stored in your vault
- **Open Source**: Fully auditable codebase
## License
This project is licensed under the MIT License - see the [LICENSE](LICENSE) file for details.
- Powered by [Anthropic Claude](https://anthropic.com), [Google Gemini](https://deepmind.google/technologies/gemini/), [OpenAI](https://openai.com), [Mistral AI](https://mistral.ai), and any OpenAI-compatible local server (LM Studio, Ollama, vLLM, etc.)
**Note**: Using a cloud provider (Claude, Gemini, OpenAI, or Mistral) requires an API key, and usage is billed by the respective provider according to their pricing — monitor your usage through provider dashboards. The Local provider requires no API key and no billing, since inference runs on your own server.
