ZeroClaw
A self-hosted AI agent for Linux, macOS and Windows. Use local models or cloud providers, connect messaging apps, and require approval for risky actions.
ZeroClaw is a self-hosted personal AI agent for people who want an assistant on their own computer or server, with control over its models and actions. It can answer messages, run shell commands, use a browser and carry out scheduled tasks. The agent runs locally.
It's a single Rust binary that runs on Linux, macOS and Windows, with Docker support. Its small runtime also suits Raspberry Pi hardware and homelab servers. ZeroClaw is open source under your choice of the MIT or Apache 2.0 license.
Local model backends include Ollama, LM Studio, llama.cpp and vLLM. With a local model, inference stays on your hardware; choosing Anthropic, OpenAI or another hosted provider sends prompts to that provider using your own key. There's no ZeroClaw cloud service between them. It also accepts OpenAI-compatible endpoints and can fall back between providers.
You can reach the agent through Telegram, Discord, Signal, Matrix, email and other channels. Schedules, webhooks and channel events can trigger tasks, while custom MCP servers extend the tools it can use. A web dashboard provides chat, memory browsing and task management. The runtime supports multiple named agents with separate isolation.
Supervised mode requires approval for medium-risk actions and blocks high-risk actions. Command allowlists, workspace boundaries and operating system sandboxes limit what the agent can access. Optional cryptographic receipts let you check that a reported tool call actually ran. Hardware tools can interact with Raspberry Pi GPIO and boards such as Arduino and ESP32.
