LocalClaw
Local-first AI agent optimized for open-source models. Privacy-focused with offline capability. Run AI completely on your own hardware.
Local-First Privacy-Focused AI Agent
LocalClaw is designed for users who prioritize privacy and want to use open-source AI models locally without cloud dependencies. Everything runs on your own hardware - your data never leaves your machine.
Philosophy: "Your data, your hardware, your control" - Complete privacy and offline capability.
Why LocalClaw?
Complete Privacy
In an era where data privacy is increasingly concerning, LocalClaw takes a different approach:
Cloud-based AI Agents:
Your Data → Internet → Cloud Server → AI Processing → Response
⚠️ Data leaves your control
LocalClaw:
Your Data → Local AI → Response
✅ Data never leaves your machineWhat This Means:
- ✅ No data sent to cloud providers
- ✅ No third-party access to your conversations
- ✅ No data retention policies to worry about
- ✅ Full compliance with privacy regulations
Offline Capability
LocalClaw works without internet:
┌─────────────────────────────────────┐
│ LocalClaw Offline Mode │
├─────────────────────────────────────┤
│ │
│ ✅ No internet required │
│ ✅ Works on air-gapped systems │
│ ✅ No cloud service dependencies │
│ ✅ Perfect for remote locations │
│ │
└─────────────────────────────────────┘Use Cases:
- Remote work locations
- Secure facilities
- Travel without reliable internet
- Cost savings on data plans
No API Costs
Run free open-source models:
| Model | Quality | VRAM Required |
|---|---|---|
| Llama 3 8B | Good | 6GB |
| Llama 3 70B | Excellent | 40GB |
| Mistral 7B | Good | 6GB |
| Qwen 14B | Very Good | 12GB |
| Phi-3 Mini | Decent | 4GB |
Cost Comparison:
| Solution | Monthly Cost (heavy use) |
|---|---|
| Cloud AI (GPT-4) | $100-500+ |
| Cloud AI (Claude) | $50-200+ |
| LocalClaw | $0 (electricity only) |
Key Features
1. Local-First Design
Everything Runs Locally:
- AI model inference
- Data storage
- Configuration
- Logs and history
Benefits:
Privacy: Your data stays on your machine
Speed: No network latency
Cost: No API fees
Control: Full control over everything
Reliability: Works without internet2. Open-Source Model Support
Supported Model Formats:
- GGUF (GGML Unified Format)
- GGML (older format)
- ONNX (Open Neural Network Exchange)
Where to Get Models:
- Hugging Face (https://huggingface.co)
- The Bloke's quantized models
- Official model repositories
Popular Models:
| Model | Size | Quality | Best For |
|---|---|---|---|
| Llama 3 8B | 4.7GB | Good | General use |
| Mistral 7B | 4.1GB | Good | Fast responses |
| Qwen 14B | 9GB | Very Good | Multilingual |
| Phi-3 Mini | 2.3GB | Decent | Low-end hardware |
3. Offline Capability
What Works Offline:
- ✅ All AI inference
- ✅ Conversation history
- ✅ File operations
- ✅ Local automations
What Requires Internet:
- ⚠️ Model downloads (one-time)
- ⚠️ Model updates
- ⚠️ Web search features (if enabled)
4. Privacy Features
Privacy by Design:
┌─────────────────────────────────────┐
│ LocalClaw Privacy Stack │
├─────────────────────────────────────┤
│ │
│ 🔒 Local Processing │
│ 🔒 Encrypted Storage (optional) │
│ 🔒 No Telemetry │
│ 🔒 No Analytics │
│ 🔒 No Data Collection │
│ │
└─────────────────────────────────────┘Installation
Prerequisites
| Requirement | Details |
|---|---|
| RAM | 8GB minimum, 16GB recommended |
| Storage | 20GB for models + data |
| GPU | Optional (speeds up inference) |
| OS | Windows 10+, macOS 12+, Linux |
Method 1: Quick Start (Recommended)
Step 1: Install
git clone https://github.com/sunkencity999/localclaw
cd localclaw
npm installStep 2: Download a Model
# Download Llama 3 8B GGUF model
# (~5GB, one-time download)
npm run download-model llama-3-8bStep 3: Configure
# config.yaml
model:
type: local
path: ./models/llama-3-8b.gguf
context_length: 4096
gpu_layers: 35 # Set to 0 for CPU-onlyStep 4: Run
npm startMethod 2: Docker
docker run -d --name localclaw \
-p 8080:8080 \
-v ./models:/app/models \
-v ./data:/app/data \
--gpus all \
localclaw/latestMethod 3: Ollama Integration
If you already use Ollama:
# LocalClaw can use existing Ollama models
# Just configure:
model:
type: ollama
ollama_url: http://localhost:11434
model: llama3Configuration
Basic Configuration
# config.yaml
# Model settings
model:
type: local
path: ./models/llama-3-8b.gguf
# Model parameters
context_length: 4096
temperature: 0.7
max_tokens: 2048
# GPU acceleration
gpu_layers: 35 # -1 for all layers on GPU
# Storage settings
storage:
data_path: ./data
encryption: false # Enable for encrypted storage
# Performance settings
performance:
threads: 8 # CPU threads for inference
batch_size: 512GPU Acceleration
NVIDIA GPU:
gpu:
enabled: true
layers: 35 # Number of layers on GPU
memory: 8GB # VRAM allocationApple Silicon (M1/M2/M3):
gpu:
enabled: true
metal: true # Use Metal APICPU Only (no GPU):
gpu:
enabled: false
threads: 8 # Use more CPU threadsUse Cases
Privacy-Critical Applications
Scenario: Handle sensitive data (legal, medical, financial)
Why LocalClaw:
Sensitive Data → LocalClaw → Processing → Response
│
└── Never leaves your machine
✅ HIPAA compliant
✅ GDPR compliant
✅ Attorney-client privilege maintainedExamples:
- Legal document analysis
- Medical record summarization
- Financial data processing
- Confidential business analysis
Offline Environments
Scenario: Work without reliable internet
Setup:
1. Download models while online
2. Copy to offline machine
3. Run LocalClaw completely offlineUse Cases:
- Remote research stations
- Maritime vessels
- Rural locations
- Secure facilities
Air-Gapped Systems
Scenario: Maximum security isolation
Implementation:
┌─────────────────────────────────────┐
│ Air-Gapped System │
│ │
│ ┌─────────────┐ │
│ │ LocalClaw │ │
│ │ + Local AI │ No network │
│ │ Model │ connection │
│ └─────────────┘ │
│ │
│ Data enters via USB only │
└─────────────────────────────────────┘System Requirements
| Component | Minimum | Recommended |
|---|---|---|
| CPU | 4 cores | 8+ cores |
| Memory | 8GB RAM | 16-32GB RAM |
| Storage | 20GB SSD | 100GB+ SSD |
| GPU | Optional | 8GB+ VRAM (NVIDIA/AMD) |
| OS | Windows 10, macOS 12, Linux | Latest |
Performance Expectations
| Hardware | Tokens/second |
|---|---|
| M3 Max | 30-50 tok/s |
| RTX 4090 | 40-60 tok/s |
| RTX 3060 | 20-30 tok/s |
| CPU Only (8 core) | 5-10 tok/s |
| CPU Only (4 core) | 2-5 tok/s |
Comparison with Alternatives
| Feature | LocalClaw | OpenClaw | Cloud AI |
|---|---|---|---|
| Privacy | ⭐⭐⭐⭐⭐ Complete | ⭐⭐⭐ Good | ⭐ Data leaves |
| Offline | ✅ Full | ⚠️ Limited | ❌ No |
| Cost | $0 API | $ API | $$ API |
| Speed | Medium | Fast | Fastest |
| Model Quality | Good | Best | Best |
| Hardware | 8GB+ RAM | 2GB RAM | Any |
Pros & Cons
Advantages
| Advantage | Explanation |
|---|---|
| Complete Privacy | Data never leaves your machine |
| Offline Capable | Works without internet |
| No API Costs | Free open-source models |
| Full Control | You control everything |
| Compliance | HIPAA, GDPR friendly |
| No Rate Limits | Use as much as you want |
Limitations
| Limitation | Explanation |
|---|---|
| Hardware Requirements | Needs 8GB+ RAM |
| Slower Than Cloud | Local inference is slower |
| Model Quality | Open-source models less capable |
| Model Management | You manage model downloads |
| Storage | Models take significant space |
Pricing
LocalClaw Software: Completely FREE (MIT License)
Costs:
- Software: Free
- Models: Free (open-source)
- Electricity: ~
$5-20/month depending on usage - Hardware: Your existing computer or one-time purchase
Savings vs Cloud AI:
Cloud AI (heavy use): `$100-500`/month
LocalClaw: `$0`/month (after hardware)
Break-even: 1-6 monthsCommunity and Support
- GitHub: https://github.com/sunkencity999/localclaw
- Issues: https://github.com/sunkencity999/localclaw/issues
- Discussions: https://github.com/sunkencity999/localclaw/discussions
License
MIT License - Free for personal and commercial use.
Summary
LocalClaw is a local-first privacy-focused AI agent offering:
- Complete Privacy -- Data never leaves your machine
- Offline Capable -- Works without internet
- No API Costs -- Free open-source models
- Full Control -- You control everything
- Compliance -- HIPAA, GDPR friendly
Best For:
- Privacy-conscious users
- Offline environments
- Sensitive data handling
- Cost-conscious heavy users
- Air-gapped systems
Not Recommended For:
- Users with low-end hardware (less than 8GB RAM)
- Those wanting fastest responses
- Users needing best model quality
- People uncomfortable with model management