Nirvana Plugin

Project Nirvana: Local-First, Privacy-First Inference

> A new way of thinking about LLM access. Your agent thinks locally, asks the cloud intelligently, and learns from the response. The cloud never sees your private data.

The Problem

Today's approach leaks your privacy and wastes 85% of your API budget.

Every time you ask your OpenClaw agent a question:

1. Your agent builds a "system prompt" containing: - Excerpts from its SOUL.md and MEMORY.md - Your personal information from its USER.md - Your entire chat history (context window) 2. All of this gets sent to cloud APIs (OpenAI, Anthropic, Google) 3. You pay for thousands of extra tokens 4. The cloud provider trains its next model on your private data

This is the current default. It's inefficient and it's a privacy disaster.

The Solution: Nirvana

Local-first inference that protects privacy and slashes costs.

Nirvana flips the paradigm:

1. Your agent thinks locally using Ollama (free, private, on your hardware) 2. For complex questions, it asks the cloud — but only sends its own carefully-crafted queries 3. Your private data never leaves your system 4. The cloud's responses are cached locally — your agent learns from them

The Paradigm Shift

| Aspect | Today (Default) | Nirvana | |--------|-----------------|---------| | Where thinking happens | Cloud only | Local first, cloud when needed | | What gets sent to cloud | Your full context + system prompts | Agent's sanitized query only | | Who learns from your data | Cloud provider | You (local agent) | | Token cost per interaction | 2,000–5,000 tokens | 50–300 tokens | | Savings | — | 85%+ token reduction | | Privacy | Leaked | Protected |

What Nirvana Does

Local Inference

Privacy Enforcement

Intelligent Routing

Learning & Caching

How It Works

User asks your agent a question
    ↓
┌─────────────────────────────────────────┐
│ Nirvana Router                          │
│ "Can qwen2.5:7b answer this locally?"  │
└─────────────────────────────────────────┘
    ↙                                   ↘
[LOCAL PATH]                        [CLOUD PATH]
80%+ of queries                     20%- of queries
Ollama (qwen2.5:7b)                 OpenAI/Anthropic/Google
Free                                Pay for answer
Private                             Cloud sees sanitized query only
~1s latency                         ~3s latency
Result cached locally               Result cached locally
    ↓                                   ↓
    └─────────────┬─────────────────────┘
                  ↓
    Agent answers your question using:
    - Local inference (primary)
    - Cloud intelligence (if needed)
    - Cached knowledge (if available)
    
    YOUR PRIVATE DATA NEVER LEFT YOUR SYSTEM

Installation

Prerequisites

Two Paths

#### Path A: Use Bundled Ollama + qwen2.5:7b (Out-of-box)

# Install the plugin
clawhub install shivaclaw/nirvana
Start Ollama container (pulls auto on first run)
docker run -d -p 11434:11434 ollama/ollama
Verify
openclaw nirvana status

#### Path B: Use Existing Local LLM (Any Provider)

# Install the skill (context stripping only)
clawhub install shivaclaw/nirvana-local
Configure endpoint
openclaw nirvana configure --local-endpoint http://your-llm:5000
Verify
openclaw nirvana status

Cost Impact

Token Savings

| Scenario | Today | With Nirvana | Savings | |----------|-------|--------------|---------| | 10 questions/day | 20,000 tokens/day | 3,000 tokens/day | 85% | | 100 questions/day | 200,000 tokens/day | 30,000 tokens/day | 85% | | Monthly cost (OpenAI GPT-4) | $500–$1,000 | $75–$150 | 85% |

Local inference is free. Only pay for the 15%–20% of queries that truly need frontier models.

Privacy Guarantee

What Never Leaves Your System

What Optionally Goes to Cloud

Privacy Audit Trail

# View what was sent to cloud this session
openclaw nirvana audit-log
Output:
2026-04-24 14:23:45 — CLOUD API CALL
Original query: [REDACTED]
Sanitized query sent: "Explain quantum entanglement"
Response cached: Yes
User data in request: None

Platform Support

| Platform | Status | Notes | |----------|--------|-------| | Linux (Ubuntu/Debian) | ✅ Full | Ollama container + native binaries | | macOS (Intel/ARM) | ✅ Full | Ollama via Docker or native | | Windows (WSL2) | ✅ Full | Ollama in WSL2 container | | VPS (Hostinger, DigitalOcean, AWS) | ✅ Full | Docker Compose ready | | Docker container | ✅ Full | Orchestrated via docker-compose | | Air-gapped (offline) | ✅ Full | Local-only mode (no cloud fallback) |

Configuration

Basic Setup

{
  "nirvana": {
    "mode": "local-first",
    "local_model": {
      "provider": "ollama",
      "endpoint": "http://ollama:11434",
      "model": "qwen2.5:7b",
      "timeout_ms": 180000
    },
    "routing": {
      "local_threshold": 0.75,
      "max_local_tokens": 8000,
      "cloud_fallback": true
    },
    "privacy": {
      "strip_soul": true,
      "strip_user": true,
      "strip_memory": true,
      "audit_logging": true
    }
  }
}

Custom Local LLM (Non-Ollama)

{
  "nirvana": {
    "local_model": {
      "provider": "custom",
      "endpoint": "http://your-llm-server:5000",
      "api_format": "openai-compatible",
      "model": "your-model-name",
      "timeout_ms": 120000
    }
  }
}

Use Cases

✅ Perfect For

⚠️ When to Use Cloud

Philosophy

Your agent should train itself. The cloud should not train on you.

Today's default paradigm:

Nirvana's paradigm:

What's Included

| Component | Purpose | |-----------|---------| | router.ts | Decides local vs cloud routing | | context-stripper.ts | Removes private data before cloud API calls | | privacy-auditor.ts | Logs all boundary crossings | | response-integrator.ts | Caches cloud responses locally | | ollama-manager.ts | Handles Ollama lifecycle + model management | | metrics-collector.ts | Tracks performance + cost + privacy | | config.schema.json | Configuration validation |

Performance

Benchmarks (qwen2.5:7b on 4-core CPU)

Optimization

Support & Community

License

MIT-0 — Free to use, modify, and redistribute. No attribution required.

*Your agent deserves privacy. Nirvana makes it real.*

Basic Setup

{
  "nirvana": {
    "mode": "local-first",
    "local_model": {
      "provider": "ollama",
      "endpoint": "http://ollama:11434",
      "model": "qwen2.5:7b",
      "timeout_ms": 180000
    },
    "routing": {
      "local_threshold": 0.75,
      "max_local_tokens": 8000,
      "cloud_fallback": true
    },
    "privacy": {
      "strip_soul": true,
      "strip_user": true,
      "strip_memory": true,
      "audit_logging": true
    }
  }
}

Custom Local LLM (Non-Ollama)

{
  "nirvana": {
    "local_model": {
      "provider": "custom",
      "endpoint": "http://your-llm-server:5000",
      "api_format": "openai-compatible",
      "model": "your-model-name",
      "timeout_ms": 120000
    }
  }
}