Public Beta v0.1
Your AI.
What it does
Real tools. Real results.
No fluff.
O.A.S.I.S. isn't a pretty chat wrapper — it's an agent with access to tools that runs local models and knows when to use them.
Multi-model support
Run powerful AI models on your own computer. No cloud, no internet required. Works with popular models like Llama, DeepSeek, Mistral, and more. Automatic hardware detection.
Real web search
When your assistant doesn't know something, it searches the web for you — automatically. No made-up answers. Gets real information, then responds.
7 built-in tools
Run commands, read and edit files, search your code, browse the web, and more — all through simple conversation. The assistant uses real tools, not tricks.
App control
Ask it to open Spotify, check what's running, or manage your windows. It controls apps on your computer — no coding required from you.
Skills + Workflows
Set up custom skills that activate when you mention a topic. The assistant detects what you're asking and runs the right steps automatically.
Voice + TTS
Talk to your assistant with a microphone. It listens, responds, and can even read replies out loud. Works completely offline.
Remembers your conversations
Your assistant remembers past chats and brings up relevant information when you need it. Like having a memory that actually works.
Thinking you can see
Watch your assistant think, search, and reason in real-time. You can peek inside its decision process or just enjoy the streaming response.
O.A.S.I.S. Pro — coming soon
Additional tools, multi-GPU support, advanced MCP integrations, priority updates, and expanded automation capabilities — for power users who need more. The free core stays complete and unlimited.
How it works
The LLM thinks, picks a tool,
and executes.
01
Pick a model
Choose from popular AI models right inside the app. It downloads automatically. O.A.S.I.S. checks your hardware and picks the best settings for you.
02
It decides what to do
When you ask a question, the assistant figures out whether it already knows the answer, needs to search the web, or should use one of its tools.
03
Uses real tools
If it needs to search, it searches the web. If it needs to read a file, it reads it. Real actions, real results — not made-up answers.
04
Responds with real info
After running its tools, it puts together a clear answer based on real data. You can see its thinking process if you want, or just read the response.
Execution flow (agentic loop)
Thinking · do I need to search?
Now
internet_search("world cup 2026")
Tool call
Synthesize results + respond
Streaming
$
oasis start --model deepseek-r1:8b
Model loaded (4.7 GB, 8 layers on GPU)
Bridge at 127.0.0.1:8765
5 skills active · 7 tools ready
→ Ready · waiting for input
Privacy first
Your prompts never
leave your machine.
No telemetry. No training on your data. No account required. O.A.S.I.S. is software you own, not a service you rent.
100% offline inference
Run any compatible model without an internet connection. Your prompts are processed entirely on your CPU or GPU — nothing is transmitted.
Local storage
Conversations, files, and settings live in a local database you control. Export, delete, or move your data whenever you want.
Zero telemetry
O.A.S.I.S. collects no usage data. No analytics, no crash reports without consent, no subscription server phoning home.
You control network access
Every tool that needs the internet — web search, remote APIs — requires explicit activation. Everything off by default.
Why go local
Your machine vs. their cloud.
You decide.
| Cloud AI OTHERS |
O.A.S.I.S. On your computer |
|
|---|---|---|
| Privacy Where do your conversations go? |
Sent to their servers. They can read, store, and train on your data. | Everything stays on your computer. No internet needed. No one else sees it. |
| Monthly cost What do you pay? |
$20–$200/month. Prices go up. Limits go down. | Free core. No subscription. No usage caps. No hidden costs. |
| Usage limits How much can you use it? |
Rate limits, message caps, "you've reached your limit" screens. | Unlimited. Chat all day, every day. It's your hardware. |
| Internet required Does it work offline? |
No connection = no AI. Offline means nothing works. | Fully offline. Airplane mode? No problem. Hotel WiFi dead? Still works. |
| You own your data Who controls your history? |
They own the servers. They can delete, move, or lose your data. You have no control. | You control everything. Export, delete, or back up your conversations whenever you want. |
| Setup What does it take to start? |
Sign up, enter credit card, configure API keys. | Download the installer. Pick a model. Chat. That's it. |
Supported models
Use any model.
Or all of them.
O.A.S.I.S. runs popular open-weight models directly on your computer. No internet needed. You can also connect to cloud APIs if you want more power.
Local
DeepSeek R1 8B
Reasoning · 8 GB RAM · Recommended
Local
Llama 3.2 8B
Fast · 8 GB RAM
Local
Mistral 7B
Balanced · 7 GB RAM
Local
Phi-3 Mini
Lightweight · 4 GB RAM
Local
Any GGUF
Llama.cpp, Qwen, Gemma, etc.
Remote
OpenAI / compatible
GPT-4o, Groq, Anthropic, DeepSeek API
Common questions
Before you download.
Hardware, privacy, how the assistant works, and what to expect.
No. Three steps and you're done:
1. Download the installer for Windows (10 or 11).
2. Pick a model from the built-in catalog — it downloads automatically to your models/ folder.
3. Start chatting. No API keys, no accounts, no cloud setup.
Everything runs on your machine. You don't need Docker, Python, or a terminal. Just the installer.
Seven local tools: bash (run commands), file (read/write files), edit (search and replace), regex (search code), web_search (DuckDuckGo), memory (persistent key-value storage), and mcp (basic MCP client).
Plus the ControllerTool lets you open, close, and control any Windows app with 5 automatic fallback levels.
When the model is unsure, the system forces it to use internet_search before answering. Three layers of protection:
1. Prompt: Dynamic instructions before every message push the model to use tools when in doubt.
2. Post-response detection: If the model generates text like "according to search results" without having actually searched, the system detects the hallucination, runs internet_search automatically, and retries.
3. Preserved history: Retries keep the full conversation context so the model learns from its mistake.
A GPU speeds things up significantly but is not required. On a modern CPU, a 7–8B model generates 5–15 tokens/second — fast enough for conversation. If you have an NVIDIA GPU, O.A.S.I.S. offloads layers automatically.
The recommended starting point is 8 GB RAM running deepseek-r1:8b in CPU mode.
No. With a local model, inference happens entirely on your machine. O.A.S.I.S. has no telemetry, no analytics, and no server that reports activity. You can use it in airplane mode permanently.
The only exception is if you explicitly enable a remote endpoint (like the OpenAI API) in Settings → Model. In that case, your prompts go to that provider — with your explicit consent, off by default.
Yes, the core is and always will be free. You provide the hardware and the models (you download them yourself). There are no subscription plans, no usage limits, and no feature restrictions on the local assistant.
We're also working on O.A.S.I.S. Pro — additional tools and integrations for power users who want more. The free version stays complete and unlimited.
A skill is a set of instructions + workflow that activates when you mention a specific topic. The IntentRouter compares each message against installed skill triggers (threshold ≥ 0.6). If there's a match, it runs the defined workflow steps.
Skills live in data/skills/ and load on startup. You can create your own.
Ollama, LM Studio, and Jan.ai are great at running models — but they're essentially chat wrappers. O.A.S.I.S. is an agentic system on top of your local model:
• Built-in tools: web search, file editing, bash, app control — the assistant can actually do things, not just chat.
• Skills & workflows: create automated routines that trigger on specific topics.
• Persistent memory: remembers your preferences, projects, and past conversations across sessions.
• Voice + TTS: fully offline speech input and output.
• Proactive mode: the assistant can suggest actions based on your usage patterns.
Currently O.A.S.I.S. is designed for Windows 10 and Windows 11. macOS and Linux support are on the roadmap — follow our GitHub for updates.
Local AI explained
How O.A.S.I.S. compares to
other local AI tools
Ollama, LM Studio, and Jan.ai run models well. O.A.S.I.S. goes further — it's an agent that acts on your machine, not just a chat window.
Agentic, not just conversational
Other local AI frontends send your message to a model and display the reply. O.A.S.I.S. decides what to do: search the web, run a command, read a file, open an app — then does it. It's an agent that uses tools, not a chat wrapper. When it doesn't know something, it searches automatically instead of guessing.
Built-in tools (7+ with fallbacks)
O.A.S.I.S. ships with bash execution, file editing (read/write/search/replace), web search, app control (5-level fallback chain), persistent memory, MCP client support, and more. These aren't plugins you need to find and install — they're built in and the assistant knows when to use each one.
Skills, workflows, and automation
Create custom skills that trigger when you mention a topic. Build visual workflows with drag-and-drop nodes. The assistant can run multi-step automated routines, save plans as reusable workflows, and connect to MCP servers for additional capabilities. No coding required.
Voice, memory, and proactive assistance
Full offline voice input (STT) and output (TTS), persistent semantic memory (ChromaDB + cross-encoder reranking), screen awareness, file watching, and a proactivity engine that can suggest actions, remind you of events, and monitor system health — all while respecting your privacy.
Start talking to your
own AI today.
Free. No account. No API key. Just you and your local model.
Requires 8 GB RAM minimum · Windows 10+ or Windows 11