Atlas - AI Local Agent

A lightweight, fully local and private AI assistant. It runs entirely on your machine (127.0.0.1), and nothing is sent to the network.

The repository contains only the code (around 150 KB). Models, the engine (Ollama), OCR languages and the offline Wikipedia are downloaded on first use, so you do not clone a huge folder.

Features

• Chat with saved conversation history (create, rename, delete)
• Local memory of facts, plus access to all past conversations
• Screen awareness via OCR, an editable knowledge graph, proactive messages and offline Wikipedia
• Ask about your own files: drop .txt, .md or .pdf in a folder and the agent answers from them (local RAG with nomic-embed-text)
• One click backup: export and restore your memory, graph and conversations
• Optional encryption at rest: protect memory, conversations, graph and reminders with a password (standard library only, no native crypto)
• Natural language reminders ("remind me of this tomorrow") that fire through the proactive channel
• System tray icon and a global hotkey (Ctrl+Alt+A) to open the agent fast
• Light and dark theme, and three interface languages (PT, EN, ES)
• Every skill can be toggled on or off at any time
• Master switch to pause all background activity (no screenshots, no learning)
• Model picker: install, switch and delete models from the interface
• Built in system monitor (VRAM, CPU, RAM)

Requirements

• Python 3.10 or newer (tested on 3.14)
• Ollama, the model engine (installable from inside the interface)
• Tesseract OCR, optional, only needed for the "see the screen" feature: https://github.com/UB-Mannheim/tesseract/wiki

Install

git clone https://github.com/pablokaua03/Atlas-AI-Local-Agent.git atlas
cd atlas
pip install -r requirements.txt

Run

• Windows: double click launcher.pyw (or create a shortcut). It has Start, Stop and Open buttons and manages Ollama for you.
• Any OS: run python server.py and open http://127.0.0.1:5005

First time

1. Open http://127.0.0.1:5005
2. In the banner, click "Install Ollama" (downloads and installs in the background).
3. In Settings, Model section, click Download on the model you want.
4. Turn on the skills you want and start chatting.

Optional

• Offline Wikipedia: in Settings, Wikipedia, click the download button. It fetches a compact .zim for your language from Kiwix and builds the index.
• OCR languages are downloaded automatically based on the interface language.

Privacy

Everything runs locally through Ollama on 127.0.0.1. No telemetry, no cloud, no account. Your data (config.json, conversas.json, memoria.json, grafo.json, observacoes.json, lembretes.json, prints, docs and docs_index.json) stays only on your machine and is listed in .gitignore.

You can also turn on encryption at rest in Settings, Security. A password is turned into a key with scrypt, and memory, conversations, graph and reminders are stored encrypted on disk (HMAC-SHA256 keystream with encrypt-then-MAC). The password is never written to disk; it is only kept in memory while the vault is unlocked. If you forget the password, the data cannot be recovered.

License

Released under the MIT License. See the LICENSE file for details.