Nullclaw
Fastest, smallest, and fully autonomous AI assistant infrastructure written in Zig
Last verified:
What is Nullclaw?
NullClaw is a high-performance, zero-dependency, single-binary execution engine for autonomous AI agents. Written entirely in Zig, it compiles into a tiny 678 KB static binary that boots in under 2 milliseconds and runs on approximately 1 MB of RAM. It features integrated AI providers, a tool execution engine, conversation memory, and a multi-channel gateway—everything needed to run agents locally or deploy them to production. NullClaw is part of the NullHub ecosystem.
Key features include a universal model multiplexer supporting 50+ providers (OpenRouter, OpenAI, Anthropic, Groq, Ollama, and more), automatic model discovery, streaming completions, and key rotation. The tool execution engine securely executes shell commands, file operations, and custom scripts with configurable approval policies and strict sandboxing. NullClaw supports 17 communication channels including Terminal, WebSocket, Telegram, Discord, Slack, Signal, WhatsApp, Matrix, iMessage, IRC, Email, Mattermost, LINE, Lark/Feishu, DingTalk, QQ, OneBot, and MaixCam. Advanced memory features include auto-compacting context windows, session save/restore across restarts, and structured memory for tool outputs.
NullClaw is ideal for three main use cases: interactive agent sessions for single-agent workflows handling complex multi-step tasks, headless worker execution for automated background pipelines without human supervision, and browser-based operation when paired with NullHub for a visual operator dashboard with real-time streaming and tool execution timelines. It targets developers, DevOps engineers, edge computing users, and anyone needing autonomous AI agents on constrained hardware like Raspberry Pi, ARM SBCs, microcontrollers, or cloud VMs.
Nullclaw pricing
Pricing model: Freemium
NullClaw is completely free and open source under the MIT license. There is no paid tier or subscription. Users need to provide their own AI API keys from providers like OpenRouter, OpenAI, Anthropic, Groq, or Ollama. The only costs are the AI API usage fees paid directly to the model provider. Homebrew installation and source builds are free. NullHub ecosystem pairing may have separate pricing.
Nullclaw pros
- Ultra-small 678 KB static binary
- Zero external dependencies
- Boots in under 2 milliseconds
- Runs on approximately 1 MB RAM
- Written entirely in Zig for safety and speed
- Supports 50+ AI providers
- 17 communication channels work correctly
- No garbage collection overhead
- Encrypted API keys with ChaCha20-Poly1305
- Automatic model discovery and selection
- Configurable autonomy levels (Supervised/Autonomous/Restricted)
- Auto-compacting context windows for long-running threads
- Works on ARM, x86, and RISC-V architectures
- Homebrew installation available
- Open source under MIT license
- Compatible with OpenClaw configurations
- Built-in file, shell, and search tools
- Streaming and non-streaming completions supported
- Key rotation and rate limit handling built-in
Nullclaw cons
- Requires Zig 0.15.2 or 0.16.0 to build from source
- Smaller plugin ecosystem compared to OpenClaw
- No polished graphical interface out of the box
- Most interaction happens through command line
- Documentation focused on core infrastructure usage
- You must provide your own AI API keys (BYOK)
- WebChannel requires manual config.json setup
- Newer project with less community maturity
Frequently asked questions about Nullclaw
What is NullClaw?
NullClaw is a high-performance, autonomous AI agent runtime written entirely in Zig. It is a zero-dependency, single-binary execution engine that compiles to a 678 KB static binary, boots in under 2 milliseconds, and runs on approximately 1 MB of RAM. It includes integrated AI providers, tool execution, memory management, and multi-channel gateway support for running autonomous AI agents locally or in production.
What AI providers does NullClaw support?
NullClaw supports 50+ AI providers including OpenRouter (recommended, aggregates 100+ models), Anthropic (Claude), OpenAI (GPT-4, o1), Groq (fast inference), Ollama (local models), and many more. The universal model multiplexer provides automatic model discovery, streaming and non-streaming completions, key rotation, and rate limit handling.
How do I install NullClaw?
You can install NullClaw via Homebrew with 'brew install nullclaw', or build from source by cloning the GitHub repository and running 'zig build -Doptimize=ReleaseSmall'. Prerequisites include Zig 0.15.2 or 0.16.0, Git, and an AI provider API key. After installation, run 'nullclaw onboard --interactive' for initial configuration.
What communication channels are supported?
NullClaw supports 17 communication channels: Terminal, WebSocket/WebChannel, Telegram, Discord, Slack, Signal, WhatsApp, Matrix, iMessage, IRC, Email, Mattermost, LINE, Lark/Feishu, DingTalk, QQ, OneBot, and MaixCam. All 17 channels work correctly as of version v2026.2.23, with support for bot tokens, allowlists, guilds, and pairing modes.
How does NullClaw handle memory and context?
NullClaw features advanced conversation history and state persistence with auto-compacting context windows for long-running, multi-turn agent threads. It supports session save/restore across restarts and structured memory for tool outputs. Memory backends include SQLite (recommended with hybrid vector + FTS5), Markdown (simple file-based), PostgreSQL (for production), and Redis (caching layer).
What security features does NullClaw have?
NullClaw is secure by design with Zig safety guarantees and strict sandboxing. API keys are encrypted with ChaCha20-Poly1305. It offers configurable autonomy levels: Supervised (requires approval for risky commands), Autonomous (full automation), and Restricted (read-only). Users can set workspace restrictions, command allowlists, and resource limits. Tool execution has robust approval policies with configurable auto-approve and deny lists.
Can I run NullClaw on embedded hardware?
Yes, NullClaw is specifically designed for edge and embedded deployment. Its 678 KB static binary fits on any $5 board, requires only libc, and runs comfortably on low-cost ARM SBCs, microcontrollers, and edge boards with around 1 MB of RAM. It is fully agnostic across ARM, x86, and RISC-V architectures with no heavyweight runtimes or virtual machines needed.
What is the difference between NullClaw and OpenClaw?
NullClaw is written in Zig with a 678 KB static binary, near-zero ~1 MB memory footprint, and instant <2 ms startup, targeting efficiency and security on edge hardware. OpenClaw is written in TypeScript/Node.js with a larger resource footprint, requiring powerful hardware, but offers a richer plugin ecosystem and more high-level integrations. NullClaw is better for efficient always-on assistants on modest servers; OpenClaw excels for power users needing feature depth on desktop-class machines.
How do I add a web chat UI to NullClaw?
To add a web chat UI, add 'channels.web' configuration to ~/.nullclaw/config.json with listen address, port (default 32123), path (/ws), and message_auth_mode set to 'pairing'. Then run 'nullclaw gateway'. In the Chat UI pairing screen, use endpoint 'ws://127.0.0.1:32123/ws' and PIN '123456' for local pairing mode. Note that 'nullclaw gateway' alone is not enough for browser pairing.
Is NullClaw free to use?
Yes, NullClaw is completely free and open source under the MIT license. There are no paid plans or subscriptions. The only cost is providing your own AI API keys from providers like OpenRouter, OpenAI, or Anthropic, and paying the API usage fees directly to those model providers. Homebrew installation and source builds are both free.