OpenClaw
A local Gateway links models and tools to chat channels, with host access requiring careful sandboxing.
What it is
OpenClaw is an open-source AI assistant that runs on a personal device or your own server. After installation and onboarding, you can send messages from the Control UI, the terminal interface, or channels such as WhatsApp, Telegram, Slack, Discord, Signal and iMessage; the Gateway receives the message, calls the configured models and tools, saves the session, and sends the result back to the original entry point.
Unlike the common assistants that work only in a standalone web page and are tied to a fixed model, OpenClaw connects models, tools, sessions, message channels and companion apps to a single Gateway. Models can come from hosted services or a local environment, and the same program can both run long-term for an individual and be deployed for a small group of people who trust one another.
How it runs
The installer supports macOS, Linux and Windows; when a suitable Node.js is missing, it also handles the runtime. The shortest path is to install the program, complete onboarding, confirm the Gateway status, and then open the control interface:
curl -fsSL https://openclaw.ai/install.sh | bash
openclaw onboard --install-daemon
openclaw gateway status
openclaw dashboardWindows can use the official PowerShell install script. If you manage Node.js yourself, you can also install the openclaw package published to npm. Repository development uses a pnpm workspace, and npm install cannot be run directly in the root directory.
Core mechanics
- Local control plane: the Gateway manages sessions, tools, events and channel connections, and the web console, CLI and TUI all work through it.
- Channel access: the same assistant can appear in everyday chat tools, so interaction does not have to be fixed in a standalone web page.
- Device capabilities: companion apps and nodes can connect voice, Canvas, camera, screen and local device operations.
- Extension system: tools, skills and plugins are used to add capabilities, and the model layer is compatible with both hosted and local sources.
Who it is for
It suits people who want the assistant to run long-term on their own computer and to be callable at any time from the chat software they use every day. If you want to maintain a shared assistant for a small group of people who trust one another, you can also keep using the same Gateway and adjust the configuration as needed.
Usage boundaries
The project uses the MIT license, while model costs depend on the model provider you connect. Tools in the main session run on the host machine by default, so external messages must be treated as untrusted input. Before connecting other users or exposing the Gateway to the public internet, you need to configure pairing, access policies and sandboxing. Team mode is intended for environments where members trust one another, and it is not equivalent to a public service that has multi-tenant isolation by default.