The Problem

Default OpenClaw transmits 2,000–5,000 tokens per query to cloud APIs. Your identity files, conversation history, memory, and sensitive context get sent unencrypted over the wire to third-party inference providers every single turn.

Even with a well-intentioned provider, this creates a massive attack surface:

• Data breaches: One API compromise exposes all your sessions
• Lateral movement: Your identity data is now a target for sophisticated adversaries
• Inference extraction: Attackers can learn your agent's personality, habits, and decision patterns
• Compliance risk: HIPAA, GDPR, SOX — if you're handling regulated data, you're violating the law by default

The Solution: Local-First with Smart Fallback

Project Nirvana keeps 80%+ of computation local on your hardware, using cloud APIs only as a stateless fallback for frontier reasoning tasks.

When fallback is necessary, the context-stripper removes:

• Identity files (SOUL.md, USER.md, AGENTS.md)
• Sensitive memory (personal notes, financial data, credentials)
• Session history (conversation context)
• Custom instructions and personality

Result: 85% token reduction + 100% privacy preservation.

What You Get

Plugin (v1.0.0)

Bundles Ollama + qwen2.5:7b with intelligent routing:

• Runs local inference by default (CPU or GPU)
• Automatically falls back to Claude Haiku for complex tasks
• Strips identity before sending to cloud
• Works offline — cloud APIs optional, not required
• Drop-in replacement for default OpenClaw

Installation:

openclaw plugins install ShivaClaw/nirvana

Skill (Lightweight Context-Stripper)

If you already have a local LLM running (Ollama, vLLM, local Claude.cpp), install the skill for context-stripping only:

• Removes identity before cloud fallback
• Minimal overhead (~50ms per query)
• Works with any OpenClaw-compatible LLM

Architecture: Local-First Router

Real-World Impact

Before Nirvana: 50+ API calls/day to OpenAI, each carrying full identity context. ~$50/month in API costs. Full session data exposed.

After Nirvana: 80%+ queries handled locally. 10–15 API calls/day (fallback only). ~$5/month in API costs. Zero identity data transmitted.

Hardware Requirement: 8GB RAM minimum (Ollama + qwen2.5:7b uses ~5GB). CPU-only inference at 3–4 tokens/sec. GPU optional but recommended for 20+ tok/sec.

Why This Matters

The premise of Project Trident is that agents should become your partners over time. They learn your preferences, your goals, your personality. That data is extremely valuable — to attackers, to competitors, to advertisers.

You should own that data. Not OpenAI. Not Google. Not Anthropic.

Nirvana is our answer: keep computation local by default, use cloud APIs only when you need their raw capability, and never send your identity to third parties.