GitHubMIT

Your private AI that knows everything

AI assistant that sees your screen, understands your work, and can run locally with Ollama or another local LLM.

Your Computer
secured
📧 emails
📄 docs
💬 chats
AI runs here
Local-only mode keeps core capture on your device

Quick answer

Private AI assistant that runs locally and knows everything on your screen. screenpipe provides desktop context to local LLMs like Ollama. No data leaves your computer.

See it in action

Cloud AI has problems

ChatGPT and Claude are powerful, but come with tradeoffs.

01

Sending screen content to OpenAI raises privacy concerns

02

Corporate policies may prohibit sharing with cloud services

03

Cloud AI doesn't know what you're working on without manual context

04

Paying for API calls every time you ask a question

05

No AI when you're offline

Local AI with desktop context

screenpipe captures your screen and feeds context to a local LLM running on your machine. Ask questions, search history, get help - all without internet.

Privacy controls

Keep screen data and conversations on your computer in local-only mode. Scope optional sync, cloud AI, exports, connectors, and team workflows before enabling them.

Works offline

No internet required. Your AI works completely offline once set up.

Desktop aware

Ask 'what's this error about?' and it can search your screen history for context.

Your choice of model

Use Ollama, LM Studio, or any OpenAI-compatible local server.

How it works

1

Install Ollama

Download Ollama and pull a model. Llama 3.2 or Mistral work well for most tasks.

# Install Ollama
curl -fsSL https://ollama.com/install.sh | sh

# Pull a model (choose one)
ollama pull llama3.2        # 3B params, fast
ollama pull mistral         # 7B params, balanced
ollama pull deepseek-r1:8b  # 8B params, good reasoning

# Verify it's running
ollama list
2

Install and configure screenpipe

Download screenpipe and point it to your Ollama instance in Settings.

# Download screenpipe from screenpi.pe
# Or install via CLI:
curl -fsSL https://screenpi.pe/install.sh | sh

# Start screenpipe
screenpipe

# In Settings → AI Provider:
# - Select "Ollama"
# - Model: llama3.2 (or your chosen model)
# - URL: http://localhost:11434
3

Ask anything

Use the screenpipe chat or integrate with your own app. The LLM has access to your screen context.

# Example: Query screenpipe API with context
curl -X POST "http://localhost:3030/chat" \
  -H "Content-Type: application/json" \
  -d '{
    "message": "What error did I see in the terminal?",
    "use_context": true
  }'

Code examples

Ollama models for different use cases

Choose based on your hardware and needs

# Fast responses (4GB RAM)
ollama pull llama3.2          # 3B, general purpose
ollama pull phi3:mini         # 3.8B, Microsoft's efficient model

# Balanced (8GB RAM)
ollama pull mistral           # 7B, great for coding
ollama pull deepseek-r1:8b    # 8B, strong reasoning

# Maximum quality (16GB+ RAM)
ollama pull llama3.1:70b      # 70B, best quality
ollama pull deepseek-r1:32b   # 32B, excellent reasoning

# Coding focused
ollama pull codellama         # Optimized for code
ollama pull deepseek-coder    # Strong at programming

Verify your setup

Check that everything is working

# Check Ollama is running
curl http://localhost:11434/api/tags

# Check screenpipe is running
curl http://localhost:3030/health

# Test a simple query
curl http://localhost:11434/api/generate -d '{
  "model": "llama3.2",
  "prompt": "Hello, how are you?",
  "stream": false
}'

Key benefits

Local-only mode available for sensitive workflows
No API costs or subscriptions
Works without internet connection
Full control over the AI model
Corporate and compliance friendly

Frequently asked questions

For smaller models (3-7B parameters), 8GB RAM is sufficient. For larger models (13B+), 16GB or more is recommended. Apple Silicon Macs and NVIDIA GPUs significantly improve performance, but CPU-only works too.

Llama 3.2 (3B) is great for fast responses on modest hardware. Mistral (7B) offers a good balance. DeepSeek-R1 (8B or 32B) excels at reasoning tasks. For coding, try CodeLlama or DeepSeek-Coder.

Modern local models are surprisingly capable for most tasks. They won't match Claude Opus on complex reasoning, but they handle everyday questions, coding help, and search tasks well. The privacy and cost benefits often outweigh the capability gap.

Yes. screenpipe supports multiple AI backends. Use local AI for sensitive work and cloud AI when you need maximum capability. You can switch in settings anytime.

Get your private AI assistant

AI power without privacy compromise.