Private

Private - Local AI Privacy Router

Route queries to Gemma4 31B on your local Ollama server. Data never leaves your local machine.

How It Works

When invoked, this skill:

1. Takes your prompt
2. Sends it to Gemma4 31B via Ollama on localhost:11434
3. Returns the response inline — Claude orchestrates, but never sees your prompt content

Usage

/private What are the financial projections for client XYZ?
/private Analyze this medical report: [paste]
/private Review these employee performance notes

Workflow

When the user invokes /private:

1. Extract the user's prompt (everything after /private)
2. Run the query tool:

bun ${PAI_DIR}/skills/Private/tools/private-query.ts ""

1. Return the Ollama response to the user verbatim — do NOT summarize or rephrase it

Available Models

| Model | Size | Best For | |-------|------|----------| | gemma4:31b (default) | 31.3B | General intelligence, reasoning, analysis | | dolphin3:8b-llama3.1-fp16 | 8.0B | Fast responses, uncensored | | dolphin3:8b | 8.0B | Fast responses, lower VRAM | | llama3.1:latest | 8.0B | General purpose |

Options

/private                           # Default: gemma4:31b
/private --model dolphin3:8b       # Use specific model
/private --system "You are a CFO"  # Custom system prompt
/private --temperature 0.3         # Lower temp for precision

Privacy Guarantees

• Data stays on your local machine — never sent to Anthropic, OpenAI, or any cloud
• No logging by third parties
• No API keys transmitted
• Suitable for HIPAA/PCI-sensitive data
• Ollama runs on localhost:11434 (not exposed externally by default)

When to Use

| Scenario | Use /private? | |----------|--------------| | Client financial data | Yes | | Medical/health information | Yes | | Employee reviews or HR data | Yes | | Proprietary business strategy | Yes | | Code with trade secrets | Yes | | General coding questions | No — use Claude | | Web research needed | No — use Claude/Research | | Latest information needed | No — Ollama has training cutoff |

Limitations

• No internet access (knowledge limited to training data)
• Slower than Claude API (~5-30s depending on prompt length)
• No tool use or file access — text in, text out
• 31B model uses significant VRAM — may slow other GPU tasks

Troubleshooting

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

# Test Gemma4 directly
curl http://localhost:11434/api/generate -d '{"model":"gemma4:31b","prompt":"hello","stream":false}'

# Check GPU usage
nvidia-smi

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: CarbeneAI
• Source: CarbeneAI/Forge
• License: MIT
• Homepage: http://carbene.ai/open-source

Install and usage instructions live in the source repository linked above.