nullclaw/nullclaw

What it solves

NullClaw is a high-performance, ultra-lightweight infrastructure for autonomous AI assistants. It eliminates the heavy runtime overhead typically associated with AI agent frameworks, allowing full-featured AI agents to run on extremely low-cost hardware, such as $5 single-board computers (SBCs) or microcontrollers, with minimal RAM and near-instant startup times.

How it works

Built with Zig, NullClaw is distributed as a single static binary with no external runtime or VM dependencies. It uses a vtable interface architecture, making every core subsystem—including AI model providers, messaging channels, memory engines, and tools—fully swappable via configuration. It includes a built-in hybrid memory system (combining vector similarity and keyword search via SQLite) and implements strict security sandboxing (using Landlock, Firejail, or Docker) to isolate agent actions.

Who it’s for

Developers and enthusiasts who want to deploy autonomous AI agents on edge devices, low-power hardware, or secure, isolated environments without the need for a heavy OS or complex dependency chains.

Highlights

• Extreme Efficiency: 678 KB binary size, ~1 MB peak RAM usage, and startup times under 2 ms on Apple Silicon.
• Broad Compatibility: Supports 50+ AI providers (including OpenRouter, Anthropic, OpenAI, and local options like Ollama) and 19 communication channels (Telegram, Discord, Slack, Nostr, etc.).
• Extensive Tooling: Includes 35+ built-in tools for shell access, file manipulation, web searching, and hardware peripheral control (GPIO, I2C, SPI).
• Security-First: Features encrypted secrets (ChaCha20-Poly1305), mandatory pairing for gateway access, and multi-layer sandboxing.
• Hardware Integration: Native support for Raspberry Pi GPIO, Arduino, and STM32/Nucleo peripherals.