Ratel
golddream-y107 downloadsAI assistant for your notes—chat with your vault, semantic search with citations, multi-step research and safe edits. Works with DeepSeek, Claude, and Ollama. Local embeddings, no telemetry.
- Overview
- Scorecard
- Updates17
The graph-native AI agent for your vault.
Chat, research, and act on your linked notes. Ask what you wrote about a topic; have Ratel pull sources into a draft. Answers cite notes you can open in one click.
Why Ratel
- Graph-native — built for linked notes, not generic chat over files
- Fusion retrieval — semantic + keyword multi-way recall, automatically pulling in linked related notes; numbered citations open in one click
- Highly extensible — built-in skills & subagents, plus MCP servers (HTTP or stdio) to plug in web search and other external tools
- Deeply customizable — swap chat / embedding / rerank models, override any prompt section, choose your MCP ecosystem
- Private & safe by default — local embeddings; network only to your configured model API and the MCP servers you add; every vault edit asks your approval first
Install
Obsidian → Settings → Community plugins → Browse → search Ratel → Install → Enable.
Requires Obsidian 1.13.0+, desktop only.
Then: open Settings → Ratel → Chat model, pick a scene preset (DeepSeek / Ollama) or custom Base → wait for the first index → click the 🦡 ribbon (or run Ratel: Ask vault).
Full walkthrough: User Guide.
What you can do
Ask with citations
“What did I write about performance tuning?” — numbered [1][2] sources, click to jump.
Multi-step work
“Pull product-planning notes into a background doc” — search, read, synthesize, write (with confirmation before edits).
Know the room
Every turn injects local time. “Summarize this note” uses the active file. Daily note path, recent edits, and outlines are first-class tools.
Remember & extend
Say “remember I prefer Tailwind…” — stored as Markdown under .ratel/memory/. Drop a SKILL.md into .ratel/skills/ to teach new workflows, or attach an MCP server (Tavily, Brave…) to give the agent web search and other external tools.
Privacy
- Index and default embeddings stay on your machine
- Network access: by default only the model API you set (DeepSeek / Claude / Ollama)
- MCP servers: only the MCP endpoints you explicitly add in settings receive requests; each tool call asks your approval by default
- No analytics, no phone-home
Docs
| Doc | Contents |
|---|---|
| User Guide | Setup, scenarios, slash commands, FAQ |
| Changelog | Full release history (ranges OK for patch bursts) |
| Architecture | Ports, agent loop, tools, workers |
Issues & ideas: GitHub Issues.