Definition:
OpenClaw is an open-source platform for running and managing artificial intelligence assistants on a self-hosted computer or server. It connects AI models to tools, files and communication channels, allowing assistants to hold conversations and carry out tasks within their configured permissions.
OpenClaw is not a language model. It is the software that coordinates the assistant: it receives messages, organises sessions, provides context and manages access to tools.
Table of contents
Origins of OpenClaw
OpenClaw was created by Peter Steinberger and developed with contributions from a community of collaborators. Its predecessors include Clawdbot and Moltbot, names used before the project adopted the name OpenClaw in January 2026.
The project evolved from connecting assistants to messaging services into a platform with agent management, persistent memory, tools and its own interfaces.
Its code is distributed under the MIT licence, which permits use, modification and redistribution under its terms. The software licence is separate from the costs and conditions of connected models or external services.
How OpenClaw works
The central component is the Gateway, a service that maintains connections and coordinates communication between interfaces, sessions and agents. It can run on a personal computer or a server.
When a request arrives, OpenClaw routes it to the relevant agent and prepares the conversation context. The model can generate a response or request a tool. If the configuration allows the tool to run, its result is returned to the agent so it can continue the task.
Several elements are involved in this process:
- Model: processes context and generates responses or tool requests.
- Agent: combines the model with instructions, context and capabilities to carry out tasks.
- Session: organises a conversation and its state.
- Tools: perform specific operations, such as consulting a source, reading a file or running a command.
- Channels and interfaces: provide ways to communicate with the assistant through the web dashboard, terminal or messaging services.
This separation makes it possible to use different models and configure agents with different roles. An AI agent may have tools that another does not, even when both run within the same installation.
Channels, tools and extensions
OpenClaw provides a web control interface and terminal access. It also supports connections to services such as Telegram, Discord, Slack and WhatsApp. The availability of each channel depends on the installed components and their configuration; many are added through official plugins.
Tools make it possible to act on connected systems. For example, consulting documentation and saving a summary requires different capabilities from those needed to modify project files.
Skills are sets of instructions that explain how and when an agent should use particular capabilities. They are usually organised around a SKILL.md file and may include supporting resources. A skill is not, by itself, a connection to a service, nor does it grant access to an account.
Plugins extend the platform with components such as channels, model providers or tools. The distinction is functional: instructions guide the work, while integrations provide the capabilities used to carry it out.
Assistant memory and context
OpenClaw can retain information in workspace files for use in subsequent conversations. This information can include user preferences, decisions and activity notes.
Persistent memory is distinct from session history. History records the progress of a conversation; memory files retain information beyond that conversation and make it available for retrieval when relevant.
Saving information is not the same as training the model. Retained data is supplied or retrieved as context, without that operation changing the language model’s parameters.
Behaviour also depends on which information is saved, how it is retrieved and what context each agent receives. The existence of memory files does not mean that all their contents are included in every response.
Installation and initial setup
OpenClaw offers options for macOS, Linux and Windows, including native Windows support. It can also be deployed on a server to keep the Gateway available without depending on a personal computer remaining switched on.
The getting started guide organises setup around these steps:
- Install the software and its prerequisites, using the method appropriate to the operating system.
- Configure model access, using a compatible connection or provider credentials.
- Complete the setup wizard, which offers quick and custom setup modes.
- Check a conversation from the web dashboard, before adding further channels or tools.
- Install the Gateway as a service, if it needs to keep running in the background.
With the command-line interface installed, openclaw gateway status checks the Gateway’s status and openclaw dashboard opens the web dashboard. Messaging channels are optional: connecting one is not required to start a conversation from that dashboard.
Local execution, data and permissions
Self-hosting OpenClaw allows control over its configuration and storage, but does not mean that all processing takes place locally. If a remote model provider is used, requests are sent to that provider. Messaging channels and external tools also play a part in how data travels.
The platform supports local and remote providers. Where the Gateway runs and where the model runs are separate decisions.
The scope of actions depends on enabled tools, available credentials and configured controls. OpenClaw provides approval mechanisms, restrictions and execution isolation, whose effects depend on how the system is deployed. Running commands directly on the host machine is not the same as running them in an isolated environment.