project-nirvana-skill（项目必杀技）

Nirvana Local: Privacy-Preserving Context Stripper

For teams that already have a local inference engine. Just strip the private data before cloud API calls. Nothing else.

What This Is

Nirvana Local is a lightweight skill for OpenClaw agents that already have a local LLM inference engine (Ollama, Llamafile, vLLM, LM Studio, etc.).

It does one thing: strips private context before cloud API calls.

• ✅ Removes SOUL.md (agent identity)
• ✅ Removes USER.md (user personal data)
• ✅ Removes MEMORY.md (agent memories)
• ✅ Removes chat history (your actual questions)
• ✅ Leaves only the task at hand
• ✅ Logs every boundary crossing (audit trail)

Why You Need This

Today: Privacy Leak

When you ask your agent a question, the full system prompt goes to cloud:

Your question: "How do I build synbio?"

System prompt sent to OpenAI:
- Agent identity (SOUL.md)
- Your personal info (USER.md)
- Agent memories (MEMORY.md)
- Full chat history
- Everything costs 2,000–5,000 extra tokens

With Nirvana Local: Protected

Your question: "How do I build synbio?"

Query to local LLM first:
- Try local inference (free)
- If local fails, ask the cloud

Cloud API call (if needed):
- Original question: [STRIPPED]
- System prompt: [STRIPPED]
- Sanitized query: "How do I build synbio?" (no context)
- Cloud never sees: SOUL, USER, MEMORY, chat history
- Cost: $0.01–$0.03 (no context overhead)

Installation

Prerequisites

• OpenClaw 2026.3.24+
• Your own local LLM running at any endpoint (Ollama, vLLM, LM Studio, etc.)

Setup (3 minutes)

# 1. Install skill
clawhub install shivaclaw/nirvana-local

# 2. Configure your local LLM endpoint
openclaw nirvana-local configure \
  --local-endpoint http://localhost:11434 \
  --local-model qwen2.5:7b

# 3. Verify
openclaw nirvana-local status

# Output:
# ✅ Local LLM: qwen2.5:7b @ localhost:11434
# ✅ Privacy audit: enabled
# ✅ Context stripper: active

How It Works

Routing Decision Logic

Agent receives your question
    ↓
Try local LLM first
(qwen2.5:7b, Mistral, Llama, whatever you have)
    ↓
┌─────────────────────────────────────────┐
│ Success?                                 │
└─────────────────────────────────────────┘
   ↙ YES (80%)           ↘ NO (20%)
   
Return local answer    Ask cloud for help
                             ↓
                    Strip private context
                    (SOUL, USER, MEMORY)
                             ↓
                    Sanitized query to cloud
                    "How do I build synbio?"
                    (no personal data)
                             ↓
                    Cache response locally
                    Agent learns
                             ↓
                    Return integrated answer

What Gets Stripped

✅ Always Removed:

• SOUL.md (agent identity)
• USER.md (personal data)
• MEMORY.md (agent memories)
• Chat history (your actual questions)
• Session context (private workstreams)

✅ What the Cloud Gets:

• Sanitized query only
• Task-specific information
• Audit trail (transparent logging)

Configuration

Basic Setup

Edit ~/.openclaw/workspace/openclaw.json:

{
  "plugins": {
    "nirvana-local": {
      "enabled": true,
      "local_llm": {
        "endpoint": "http://localhost:11434",
        "model": "qwen2.5:7b",
        "timeout_ms": 180000,
        "api_format": "openai-compatible"
      },
      "privacy": {
        "strip_soul": true,
        "strip_user": true,
        "strip_memory": true,
        "strip_chat_history": true,
        "audit_logging": true
      },
      "routing": {
        "local_threshold": 0.75,
        "max_local_context_tokens": 8000,
        "cloud_fallback": true
      }
    }
  }
}

Custom API Format

If your local LLM uses a different API:

{
  "plugins": {
    "nirvana-local": {
      "local_llm": {
        "endpoint": "http://your-server:5000",
        "model": "your-model",
        "api_format": "custom",
        "custom_api_handler": "llamafile"  // or "vllm", "lm-studio", etc.
      }
    }
  }
}

Privacy Audit Trail

View What Gets Stripped

# See every boundary crossing
openclaw nirvana-local audit-log --tail 20

# Output:
# [2026-04-24 14:23:45] LOCAL HANDLING
# Question: "What's my salary range for synbio roles?"
# Handled by: qwen2.5:7b locally
# Private data: None exposed
# Cost: $0

# [2026-04-24 14:25:12] CLOUD FALLBACK (WITH STRIPPING)
# Original question: [STRIPPED]
# Sanitized query sent: "What are typical salary ranges in synthetic biology?"
# Private data stripped: SOUL.md, USER.md, chat history
# Cost: $0.02

Transparency

Every cloud API call is logged with:

• What was stripped
• What was sent
• What was cached
• Cost incurred
• Privacy boundary verified

Supported Local LLMs

Philosophy

You own the learning. The cloud provides intelligence.

Without Nirvana Local:

• Cloud provider learns from your private data every time you ask a question
• You train their next model
• Your personal information becomes their training corpus
• You pay for the privilege

With Nirvana Local:

• Your local agent learns from cloud responses
• Your private data never leaves your system
• Cloud provider learns nothing about you
• You own all the knowledge

Cost Savings

Example: 10 Questions/Day

Today (without Nirvana Local):

• 2,000 tokens/question (full context sent)
• 20,000 tokens/day
• $0.60/day (OpenAI GPT-4)
• $18/month
• Privacy: Compromised

With Nirvana Local:

• 80% local (free, private)
• 20% cloud (sanitized, no context overhead)
• 300 tokens/question average
• 3,000 tokens/day
• $0.09/day
• $2.70/month
• Privacy: Protected

Savings: $15.30/month + 100% privacy protection

When to Use

✅ Perfect For

• Agents with local LLMs already running
• Privacy-critical deployments (code, healthcare, legal, finance)
• Cost-conscious teams (85% savings)
• Air-gapped environments (local + selective cloud)

⚠️ When to Use Full Plugin

• Need automated Ollama + model setup
• No local LLM currently available
• Want out-of-box simplicity

Support

• GitHub: ShivaClaw/nirvana-skill
• Issues: GitHub Issues
• Discussions: GitHub Discussions

License

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

*Your privacy is yours to keep. Nirvana Local makes it happen.*