Manex Brain

A local AI brain for your Obsidian vault. Indexes all your notes and answers questions privately using a local Apple Silicon MLX model — no cloud, no API keys, no data leaves your machine.

Requirements

• macOS with Apple Silicon (M1 or later)
• Python 3 (via Homebrew: brew install python)
• Obsidian 1.5.0 or later

How It Works

On first load, the plugin:

1. Creates a Python virtual environment at ~/.obsidian-study-room/venv
2. Installs mlx-lm into the venv (~2 min on first run)
3. Downloads and starts mlx-community/Qwen3-4B-4bit (~2.3 GB, downloaded once)
4. Indexes all your vault notes in the background for vault-wide semantic search

After that, the server starts automatically every time Obsidian opens.

Features

• Fully local — inference runs on-device via Apple's MLX framework; nothing is sent to the cloud
• Vault-wide RAG — all notes are indexed; relevant chunks are automatically included in every answer
• Active note context — the currently open note is semantically embedded and prioritised
• Graph-aware — linked notes, backlinks, and shared-tag notes are included as context
• Clickable sources — note links in answers open the corresponding note in Obsidian
• Chat memory — conversation history is maintained within a session
• Automatic re-indexing — notes are re-indexed on create, edit, rename, or delete

Usage

1. Open Obsidian — the MLX server starts automatically in the background
2. Click the brain icon in the ribbon, or run Open Manex Brain panel from the command palette
3. Open any note and start asking questions — the plugin automatically uses it as context
4. Click any note link in an answer to open that note

Commands

Settings

Privacy

All processing is local. The plugin:

• Makes no network requests except to http://localhost:8080 (the local MLX server)
• Stores the vault index at ~/.obsidian-study-room/vault-index.json (text only, no embeddings)
• Does not collect, transmit, or store any user data

Troubleshooting

"MLX server not available" — The model may still be downloading or loading. Wait 2–5 minutes on first run; the panel will update automatically once the server is ready.

Python not found — Install Python via Homebrew: brew install python

Slow first response — Normal. The model (~2.3 GB) is downloaded and loaded into memory once; subsequent responses are fast.

Support

Visit manex.app to learn more or support development.