# Offline-First Mode Quickstart Guide 🔌
> **Version**: V14.0.0+
> **Prerequisites**: Python 3.10+, 8GB+ RAM (16GB recommended)
Boring-Gemini V14.0 introduces a true **Offline-First** architecture. This guide helps you set up a fully autonomous local development environment with zero internet dependency.
---
## 1. Quick Setup
### Step 1: Install Dependencies
Offline mode requires `llama-cpp-python` for local inference.
```bash
# Install with local support extras
pip install boring-aicoding[local]
# Or manually
pip install llama-cpp-python
```
> **GPU Acceleration**: If you have an NVIDIA GPU, install with CUDA support:
> `CMAKE_ARGS="-DGGML_CUDA=on" pip install llama-cpp-python`
### Step 2: Download a Model
Use the built-in CLI to download a recommended GGUF model.
```bash
# List recommended models
boring model list
# Download a balanced model (e.g., Llama-3-8B-Quantized)
boring model download --name "llama-3-8b-instruct-q4_k_m.gguf"
```
Models are stored in `~/.boring/models/`.
### Step 3: Enable Offline Mode
You can enable offline mode globally or per session.
**Option A: CLI Toggle (Persistent)**
```bash
boring offline enable
```
**Option B: Environment Variable (Temporary)**
```bash
export BORING_OFFLINE_MODE=true
boring start
```
---
## 2. Verification
Run the doctor command to verify your offline status.
```bash
boring doctor
```
Output should show:
```
5. Offline Mode
- Status: ENABLED
6. Local LLM Models
- Models: 1 available
- llama-3-8b-instruct-q4_k_m.gguf
```
---
## 3. How it Works
When Offline Mode is active:
1. **Network Cutoff**: All external API calls (Gemini, OpenAI, Anthropic) are blocked.
2. **Local Inference**: The Agent automatically routes LLM requests to your local GGUF model.
3. **Local Tools**: Only local tools are loaded (File Ops, Local RAG, Shell). Web search tools are disabled.
4. **Local RAG**: Queries use `SentenceTransformers` (local embeddings) and `ChromaDB` (local vector store).
### Fallback Behavior
If no local model is loaded but Offline Mode is ON, the system will error out gracefully suggesting you to run `boring model download`.
---
## 4. Performance Tuning
Create a `.env` file in your project to tune performance:
```ini
# .env
BORING_LOCAL_MODEL_PATH=~/.boring/models/my-custom-model.gguf
BORING_LOCAL_CTX_WINDOW=8192
BORING_LOCAL_GPU_LAYERS=35 # Offload layers to GPU
```
---
*Last updated: V14.0.0*
