Outpost - Offline AI Station

Run LLMs locally. No internet. No limits. No cost per token.

Outpost is a self-hosted AI workstation: a native desktop app that runs language models on your own hardware, answers questions about your own documents, and never sends a byte outside your machine.

---

Features

• 🖥️ Native desktop app - built with Tauri, not a browser tab
• 🧠 Multi-model support - download and swap between any Ollama model
• 📄 Document Q&A (RAG) - upload PDFs, CSVs, Markdown, DOCX → ask questions about them
• ⌨️ Full CLI - everything the app does, scriptable from the terminal
• 🔌 OpenAI-compatible API - point existing tools at localhost instead of api.openai.com
• 🔒 100% offline after setup - no telemetry, no accounts, no rate limits

---

Requirements

Tool	Version	Notes
Ollama	latest	Runs the models
NVIDIA GPU	optional	Recommended for faster inference; available VRAM determines which model sizes can be run comfortably
Windows / macOS / Linux	-	Native builds for all three

The desktop app ships with the backend bundled in - no Python install required for end users.

---

Install (Not available yet)

1. Download the installer for your platform from the Releases page
2. Install Ollama if you haven't already
3. Launch Outpost

That's it. The app starts its own backend automatically.

---

First Run

1. Open Models tab → download a model (start with llama3.2:3b, ~2GB)
2. Also download nomic-embed-text if you want document Q&A
3. Go to Chat and start talking
4. Go to Documents to upload files for RAG

---

Recommended Models

Model	Size	VRAM	Best for
llama3.2:3b	2 GB	~3 GB	Fast everyday tasks
llama3.1:8b	4.7 GB	~6 GB	General purpose
deepseek-coder-v2:16b	9 GB	~10 GB*	Coding
qwen2.5:7b	4.4 GB	~6 GB	Multilingual
nomic-embed-text	0.3 GB	~1 GB	Required for document Q&A

*Partial CPU offload on 8GB cards.

---

Architecture

┌─────────────────────────────────┐
│     Outpost (Tauri + React)     │
└────────────────┬────────────────┘
                 │  localhost:8765
┌────────────────▼────────────────┐
│     Backend (bundled, no        │
│     Python required)            │
│     ├── /chat        (SSE)      │
│     ├── /v1/*   (OpenAI compat) │
│     ├── /models                 │
│     └── /documents   (RAG)      │
└───────┬─────────────────┬───────┘
        │                 │
┌───────▼──────┐   ┌──────▼───────┐
│    Ollama    │   │   ChromaDB   │
│   :11434     │   │   (local)    │
└──────────────┘   └──────────────┘

---

Using the OpenAI Compatible API

Point any existing tool at Outpost instead of OpenAI:

from openai import OpenAI
client = OpenAI(base_url="http://localhost:8765/v1", api_key="local")

response = client.chat.completions.create(
    model="llama3.1:8b",
    messages=[{"role": "user", "content": "Hello"}]
)

Works with Continue.dev, LangChain, and most OpenAI SDK-based tools.

---

Configuration

Settings live at ~/.outpost/config.json and are also editable from the Settings tab in the app:

{
  "ollama_host": "localhost",
  "ollama_port": 11434,
  "backend_port": 8765,
  "context_length": 4096,
  "temperature": 0.7,
  "rag_enabled": true,
  "rag_top_k": 5,
  "system_prompt": "You are a helpful AI assistant running locally."
}

Using the CLI - Under development!

The CLI talks to the same backend the app uses - they share everything.

outpost status                  # check backend, models, docs
outpost models list             # see installed models
outpost models pull llama3.2:3b # download a model
outpost chat                    # interactive terminal chat
outpost docs add ./notes.pdf    # index a document
outpost docs list               # see indexed documents

---

Want to build it from source?

See DEVELOPMENT.md for the full build pipeline, project structure, and debugging guide.

---

License

MIT