NullClaw
nullclaw/nullclaw
The ultimate minimalist's AI agent framework — a 678 KB static binary that boots in under 2ms and sips just ~1 MB of RAM. It's what happens when someone dares to ask 'how small can an AI assistant actually be?' and answers with 'yes.'
Why choose NullClaw over OpenClaw?
Quick recommendation layer first, deeper analysis second. Use this before diving into metrics and architecture details.
- Safer default posture than OpenClaw for security-conscious deployments.
- Runs far leaner than OpenClaw on constrained hardware and low-cost hosts.
- Emphasizes isolation and containment where OpenClaw often prioritizes raw flexibility.
- Efficiency usually comes with narrower scope, fewer integrations, or rougher ergonomics.
- Security-sensitive self-hosters
- Edge devices and lightweight deployments
- You care more about broad integrations than minimal footprint
Limited evidence available. Use the primary sources before making a production decision.
AI decision layer last reviewed Jun 1, 2026. Helpful, but still inference-heavy enough to double-check primary sources.
Source window: GitHub metadata, README, recent commits, latest release, Reddit, Brave search
Community Pulse
Security Radar
How it's evaluated
Isolation from host OS. 10 = Fully virtualized (Docker/Wasm); 1 = Direct local execution.
Safety of external connections. 10 = End-to-end encrypted/Scoped; 1 = Plaintext/Broad access.
Traffic control. 10 = Air-gapped/Offline-first; 1 = Unrestricted internet access.
Privacy level. 10 = Zero telemetry/Zero tracking; 1 = Extensive logging/reporting.
Command safety. 10 = No unsupervised shell; 1 = Raw, unmonitored shell access.
Security radar summary for NullClaw.
- NullClaw: Sandboxing 9 of 10, API Security 8 of 10, Network Isolation 8 of 10, Telemetry Safety 9 of 10, Shell Protection 7 of 10.
Evaluation Scale: 10 = Maximum Safety / 1 = High Risk
Star Growth (2026)
Star history summary.
- nullclaw: 149 recorded points. From 155 stars on 2026-01-01 to 7,655 on 2026-06-05.
ClawVerse News
Latest articles and global buzz
Trending Mentions
Technical Showdowns
NullClaw is a fully autonomous AI assistant infrastructure implemented entirely in Zig, delivering an impossibly compact 678 KB static binary that challenges every assumption about what AI tooling requires. Unlike OpenClaw's TypeScript-based ~28 MB distribution or ZeroClaw's Rust implementation, NullClaw achieves near-zero overhead through Zig's compile-time optimizations and lack of runtime garbage collection. The framework boots in under 2 milliseconds on Apple Silicon and under 8ms on constrained 0.8 GHz edge hardware, making it uniquely suited for deployment on $5 ARM boards, microcontrollers, and embedded systems.
The architecture emphasizes complete swappability through vtable interfaces for providers, channels, tools, memory systems, tunnels, and peripherals — supporting 50+ providers, 19 channels, and 30+ tools out of the box. Security is foundational rather than bolted on, featuring multi-layer sandboxing (landlock, firejail, bubblewrap, Docker), explicit allowlists, workspace scoping, and encrypted secrets. The hybrid vector + FTS5 memory system and MCP protocol support ensure feature parity with heavier alternatives while maintaining the ~1 MB RAM footprint that defines the project's identity.
Where OpenClaw targets desktop and cloud deployments with its >1 GB memory requirements, NullClaw carves out the edge and embedded niche — running comfortably on hardware that would choke other frameworks. Recent commits show active development around Telegram integration, subagent memory management, and the innovative Hashline tool for precise, verifiable code editing. The project represents a philosophical statement about software efficiency: that AI assistants need not be resource hogs.