Ask your vault.
Get answers it can actually back up.
A local-first AI research assistant for your Obsidian vault. It sticks to what your notes actually say and fabricates nothing. Runs entirely on your machine, reuses your Claude login, no API key.
Five steps, zero hallucination
Ask
Type a question in plain language and hit Ask.
Retrieve
BM25 RAG pulls only the most relevant slices of your vault.
Ground
Claude answers only from what it found, with sources.
Humanize
A de-AI pass strips robotic phrasing so the answer reads clean and natural.
Refine
Edit answers in place or refine with follow-ups.
Everything, running on your machine
Grounded Q&A
Ask mode reads the relevant files and answers only from what your notes actually say. No fabrication, and answers are editable in place.
Multi-tab sessions
Spawn unlimited tabs. Each one is an independent session with its own auto-generated name and accent color, runs concurrently with the rest, and survives restarts.
Job-application mode
Paste a job description and question; get a grounded, first-person, humanized draft, then refine it with natural-language follow-ups.
Humanizer skill
A de-AI rewrite pass, on by default, strips robotic LLM patterns so generated text reads like a real person wrote it.
Zero-dependency RAG
A per-tab BM25 retriever pulls only the most relevant slices of your vault. Pure TypeScript, no embeddings service, far fewer tokens.
Free local web research
Optional SearXNG integration gives every agent a bounded web_search / web_read tool, with Readability parsing and a headless-Chromium fallback. No paid API.
Pluggable engines
Switch engines per tab or globally: Claude Code and Codex, plus Gemini Antigravity, OpenCode, Cursor, and Copilot.
Extensible skills
Drop a SKILL.md into a folder and it is auto-detected and embedded into every engine's prompt. No code changes required.
Polished design system
OKLCH color tokens, dark/light themes, compact density, split view, status pills, and a playful fun-mode with animated backgrounds.
Local, private, transparent
Everything runs on localhost. A Logs page shows the full activity history, including tool calls, timing, and which engine handled each request, so you can audit exactly what happened.
Research your knowledge base
Type a question, the agent reads the relevant files, and answers strictly from what it finds. Answers are editable in place.
Draft grounded applications
Paste a job description and question, get a first-person, humanized draft from your notes, then refine it with follow-ups like "make it shorter."
Switch engines per tab or globally
- Claude Agent SDK
- Session-based follow-ups
- Reuses your existing login
- OpenAI Codex CLI engine
- Same grounded workflow
- Toggle any time
A UI you'll actually want to live in
OKLCH tokens, dark & light themes, compact density, split view, and a playful fun-mode. Try the controls.