What is this?
ZeroClaw is a personal AI assistant runtime: a single Rust binary you configure and run yourself, which talks to LLM providers, listens on dozens of channels, and acts through tools like shell, browser and HTTP.
ZeroClaw is a personal AI assistant runtime: a single Rust binary you configure and run yourself, which talks to LLM providers, listens on dozens of channels, and acts through tools like shell, browser and HTTP.Is this an app or a runtime I configure myself?It is not a hosted app — it is an agent runtime binary you deploy and configure with a TOML file, wiring up providers, channels and tools yourself.
It is not a hosted app — it is an agent runtime binary you deploy and configure with a TOML file, wiring up providers, channels and tools yourself.How do I get it running?On Unix, run the install.sh script (install.sh --help shows all flags); on Windows, use the prebuilt PowerShell path from the Windows setup guide, which installs the release, updates PATH, and runs Quickstart.
On Unix, run the install.sh script (install.sh --help shows all flags); on Windows, use the prebuilt PowerShell path from the Windows setup guide, which installs the release, updates PATH, and runs Quickstart.Which LLM providers does it need an account with?It talks to Anthropic, OpenAI, local Ollama, or about 20 other providers (including any OpenAI-compatible endpoint), configured in a [providers.models.*] block in config.toml.
It talks to Anthropic, OpenAI, local Ollama, or about 20 other providers (including any OpenAI-compatible endpoint), configured in a [providers.models.*] block in config.toml.What's the catch with the config file?A working config needs at least four TOML section headers (a provider entry, an agent referencing it, and a risk profile it gates against); if a provider's streaming fails, ZeroClaw silently retries non-streaming.
A working config needs at least four TOML section headers (a provider entry, an agent referencing it, and a risk profile it gates against); if a provider's streaming fails, ZeroClaw silently retries non-streaming.Which operating systems does it support?Linux, macOS, Windows, FreeBSD, NixOS, and Docker all have platform-specific install notes.
Linux, macOS, Windows, FreeBSD, NixOS, and Docker all have platform-specific install notes.Where does my data and keys actually live?Everything runs on your own machine with your own API keys and workspace; nothing is routed through a vendor-hosted service.
Everything runs on your own machine with your own API keys and workspace; nothing is routed through a vendor-hosted service.What stops it from running dangerous commands on my machine?By default autonomy is supervised: medium-risk operations require approval and high-risk ones are blocked, backed by workspace boundaries, command policy and OS-level sandboxes (Landlock, Bubblewrap, Seatbelt, Docker). A YOLO mode exists to skip these gates on trusted dev boxes.
By default autonomy is supervised: medium-risk operations require approval and high-risk ones are blocked, backed by workspace boundaries, command policy and OS-level sandboxes (Landlock, Bubblewrap, Seatbelt, Docker). A YOLO mode exists to skip these gates on trusted dev boxes.Can it control physical hardware, not just chat?Yes — through the Peripheral trait it can drive GPIO, I2C, SPI and USB on devices like Raspberry Pi, STM32, Arduino and ESP32.
Yes — through the Peripheral trait it can drive GPIO, I2C, SPI and USB on devices like Raspberry Pi, STM32, Arduino and ESP32.Can it hook into my IDE?Yes, via ACP (Agent Client Protocol), a JSON-RPC 2.0 over stdio integration for editors and IDEs.
Yes, via ACP (Agent Client Protocol), a JSON-RPC 2.0 over stdio integration for editors and IDEs.What is the SOP engine for?It runs event-triggered Standard Operating Procedures (fired by MQTT, webhook, cron or hardware peripheral events) with approval gates and resumable runs, letting the agent carry out multi-step automated workflows.
It runs event-triggered Standard Operating Procedures (fired by MQTT, webhook, cron or hardware peripheral events) with approval gates and resumable runs, letting the agent carry out multi-step automated workflows.