ZeroClaw
Runtime OS for Agentic Workflows

ZeroClaw is a runtime operating system for agentic workflows — a lightweight, trait-driven infrastructure for building and deploying autonomous AI agents. Written entirely in Rust, it abstracts models, tools, memory, and execution so agents can be deployed anywhere with minimal overhead. It runs on hardware as small as $10 boards with less than 5MB RAM.

There are multiple ways to install ZeroClaw: (1) Homebrew: `brew install zeroclaw` (macOS/Linux), (2) One-click bootstrap: clone the repo and run `./bootstrap.sh`, (3) Pre-built binaries for Linux (x86_64, aarch64, armv7), macOS (x86_64, aarch64), and Windows (x86_64) from GitHub Releases, or (4) Build from source with `cargo build --release --locked`.

ZeroClaw and OpenClaw are both AI agent frameworks, but they differ fundamentally. ZeroClaw is a Rust-based runtime OS using less than 5MB RAM with sub-10ms cold starts, while OpenClaw is TypeScript-based using over 1GB RAM with 500s+ startup times. ZeroClaw uses a trait-driven architecture where every subsystem (providers, channels, tools, memory, tunnels) is swappable. It is designed for edge hardware and low-cost deployment.

ZeroClaw supports multiple AI providers out of the box with no vendor lock-in. It is compatible with OpenAI, Anthropic (Claude), OpenRouter, Ollama (for local models), and any custom OpenAI-compatible endpoint. Providers are implemented as swappable traits, so you can easily switch between cloud APIs and self-hosted models like Llama, Mistral, or Qwen without changing your agent code.

Yes, ZeroClaw is completely free and open-source under a dual Apache-2.0 OR MIT license. You can use it for personal, commercial, or enterprise projects without any licensing fees.

ZeroClaw is built with a security-first architecture. It features gateway pairing with device codes, deny-by-default channel allowlists, filesystem sandboxing with workspace scoping, rate limiting, and encrypted credential storage. For Docker deployments, it runs in a fully sandboxed environment. Every external access requires explicit allowlisting, making it suitable for production and enterprise use cases.

Yes, ZeroClaw supports Docker as a sandboxed runtime environment. You can run the agent, gateway, or daemon inside Docker containers for isolation and easy deployment. Docker is one of the two supported runtimes (Native and Docker). This makes it straightforward to deploy ZeroClaw in cloud environments, Kubernetes clusters, or any containerized infrastructure.

ZeroClaw supports 15+ messaging channels including CLI, Telegram, Discord, Slack, Mattermost, Matrix, Signal, iMessage, WhatsApp, Lark, DingTalk, Nostr, Email, IRC, Webhook, QQ, and Linq. Each channel implements a common trait for easy integration.

After installing ZeroClaw, you can start an interactive agent session with `zeroclaw agent`, or send a single message with `zeroclaw agent -m "your message"`. For always-on deployment, use `zeroclaw daemon` to run in the background and `zeroclaw gateway` to expose it via a secure gateway. ZeroClaw's trait-driven architecture lets you configure providers, channels, tools, and memory backends to build agents tailored to your specific use case.

ZeroClaw includes a custom full-stack search engine with zero external dependencies — no Pinecone or Elasticsearch required. It uses a hybrid approach combining vector storage (embeddings as SQLite BLOB with cosine similarity), keyword search (FTS5 with BM25 scoring), and weighted merge of both. It also supports PostgreSQL, Markdown files, or no memory backend.