Your private AI notebook.

Notes are for writing and thinking — perfect for essays, research, and knowledge bases with full AI assistance. Notebooks are for code — run Python, JavaScript, and TypeScript with instant execution and inline output. Both formats work together, and Memory Links connects them automatically.

Canto runs a native Python environment locally on your Mac — no WebAssembly, no cloud. Popular packages like numpy, pandas, and matplotlib install automatically when you import them. Variables are shared across Python, JavaScript, and TypeScript cells in the same notebook.

Memory Links is Canto’s semantic linking system that automatically finds related notes as you write. Unlike manual tagging or keyword search, it uses AI embeddings to understand meaning — surfacing relevant insights from thousands of notes instantly, including code examples from your notebooks.

Yes. After the first model download, Canto works completely offline. On planes, trains, or anywhere without internet — unlimited AI assistance, Memory Links, and semantic search are always available.

Canto ships with 10 built-in local models: Qwen 3.5 2B (1.3 GB), Qwen 3.5 4B (2.9 GB), Nemotron 3 Nano 4B (3.0 GB), Qwen 3.5 9B (5.9 GB), GPT-OSS 20B (12.0 GB), Qwen 3.6 27B (17.5 GB), GLM-4.7 Flash (18.3 GB), Qwen 3.6 35B A3B (21.4 GB), Qwen 3.5 122B A10B (75.0 GB), and Nemotron 3 Super 120B A12B (87.0 GB). All support tool calling and run locally via Metal GPU acceleration. You can also connect Ollama or LM Studio as external endpoints to use any additional model you’ve downloaded — including cloud-hosted providers like OpenAI and Anthropic through Ollama.

Absolutely. Your notes, code, and Memory Links embeddings are stored locally in an AES-256 encrypted SQLite database. When you use iCloud Drive Sync, Canto writes encrypted sync data to your own iCloud Drive folder using a Sync Passphrase you choose. Canto does not upload readable note content to a Canto server.

macOS 14 (Sonoma) or later. Requires Apple Silicon (M1 or later) for Metal GPU inference. 8 GB RAM is the minimum — enough for Qwen 3.5 2B (1.3 GB). Enable Low Memory Mode in Settings to attempt larger models on smaller machines (may reduce context or cause instability). Recommended: 16 GB+ for models up to 9B, 32 GB+ for Qwen 3.6 27B and 35B-class models, 96 GB+ for the 122B and 120B models. Models download once and cache locally.

Yes. Canto ships a built-in MCP Server that lets Claude Desktop, Cursor, Claude Code, Windsurf, OpenClaw, and any other Model Context Protocol client read and write your vault locally. Toggle it on in Settings → MCP Server. The server only accepts connections from your own Mac, an optional bearer token adds a second lock, and tools are split into always-on vault tools plus a separate opt-in automation dispatcher for scripted workflows. Per-tool approvals from your client still apply, so you stay in the loop.

Yes. Canto includes a Safe Mode launch path designed for tight machines. If free memory is low at boot, or if the previous launch crashed before the window finished loading, Canto starts in a stripped-down mode where your notes, editor, search, and UI still work normally. AI extras are deferred until you say go, and Canto can suggest a smaller model if your last one was too large.

The Memory dashboard and status bar aggregate Canto’s related processes, including heavier helpers like vision workers. The Memory modal can also break the total down by process so you can see which subsystem is using the most RAM.

Your Canto license can be activated on 1 device. To view activated devices, deactivate old machines, or transfer your license, create a LonelyDuck account using the same email from your purchase receipt. Visit your Account Dashboard to manage all your licenses in one place.