Atlas
Local-first · Open source

The AI assistant that actually remembers you

Atlas runs on your machine, reads your Obsidian vault, and reaches you by chat, voice, or WhatsApp — then does things about it.

atlas · local · localhost:5173
Connectors
os-tools
vault
whatsapp
web-search
drive
hey atlas — when's the Nightingale draft due?
search_vault · Journal, projects/nightingale.md
Thursday. Want me to text Sam a reminder?
yes, do it
Ask Atlas, or say "hey atlas"…
Audit
audit: search_vault
audit: whatsapp.send
audit: save_note
Obsidian vaultWhatsAppVoiceMCP toolsModel failoverMIT license Obsidian vaultWhatsAppVoiceMCP toolsModel failoverMIT license
Memory

Atlas remembers everything in your vault

Every conversation is journaled as Markdown. BM25 retrieval pulls the relevant notes into each turn, and Atlas auto-links what it writes back into a real wikilink graph — no separate database to manage.

Graph of Atlas vault notes linked together
your vault, as a living wikilink graph
Action

When you need something done, Atlas acts

Every capability is an MCP server — Atlas can search the web, write files, message you on WhatsApp, or scaffold a brand-new tool when nothing else fits. Every action lands in a full audit log.

whatsapp.send · to=+1•••4021
"Nightingale draft due Thursday"
audit: whatsapp.send | status=ok
audit: save_note | type=lesson
audit: cron.schedule | job=reminder
Reach

Talk to it anywhere you already are

Say "hey atlas" and talk out loud, open the local chat UI, or just text it on WhatsApp. Drop in a photo, PDF, or voice note and Atlas reads it in context.

Voice
Chat UI
WhatsApp
Capabilities

Everything Atlas ships with

Bundled connectors and behaviors, all running locally by default.

Local voice loop

"Hey atlas" wakes it up; speech-to-text runs on your machine.

Learn more

Obsidian-vault memory

Every note becomes memory via BM25 retrieval and wikilinks.

Learn more

WhatsApp built in

Linked once via QR code; the bridge starts with Atlas.

Learn more

Multi-provider failover

Ordered Model Hierarchy fails over mid-message on errors.

Learn more

Self-extending tools

Writes, validates, and enables its own MCP servers on demand.

Learn more

Multimodal

Images, PDFs, and voice notes, read in context — in-app or WhatsApp.

Learn more
0%
Local-optional — with a local model, nothing leaves your machine
MIT
Fully open source, free to run and modify
Py 3.11+
One setup script installs everything, including WhatsApp
Try it now

Interactive preview, right in your browser

Simulated tools, sample vault — a taste of Atlas before you install it.

Launch live preview
Simulated tools · sample notes · not full Atlas
Connectors
os-tools
web-search
whatsapp (sim)
When is Nightingale due?
Search local-first AI
Remember I like concise replies
Type a message…Send
Journal (sample)
2026-07-06 — Draft outline
2026-07-05 — Vault sync test
8 messages per sessionDownload full Atlas

Frequently asked questions

Is the browser preview full Atlas? No — see what's simulated →
Does my data leave my machine?+
Not with a local model backend like Ollama or LM Studio — everything, including speech-to-text, runs locally. Using a hosted API means only the relevant text is sent to that provider.
Do I need Obsidian?+
No — the vault is just a folder of Markdown files. Obsidian is a nice way to browse and edit it, but entirely optional.
What happens if a model provider goes down?+
Atlas fails over to the next tier in your configured Model Hierarchy mid-message, and tells you it switched.
Is WhatsApp required?+
No, it's on by default during setup but fully skippable, and re-run setup anytime to add it later.

Run an AI that actually remembers you