What is OpenClaw?
OpenClaw is an open-source AI agent platform that lets you build and manage intelligent conversational agents across more than 50 communication channels. With over 5,700 built-in skills, OpenClaw gives you a production-ready foundation for creating agents that can answer questions, automate tasks, and integrate with the tools your team already uses.
Whether you want a customer-support bot on Slack, a personal assistant on Telegram, or an internal knowledge agent accessible through a web UI, OpenClaw provides a single platform to make it happen.
Architecture overview
OpenClaw is built around four core subsystems that work together to receive a user message, process it intelligently, and return a useful response.
Agent Core
The Agent Core is the brain of the platform. It manages conversation state, orchestrates calls to language models, decides which skills to invoke, and maintains memory across interactions. Every message flows through the Agent Core before anything else happens.
Channel Adapters
Channel Adapters are lightweight connectors that translate between OpenClaw's internal message format and the APIs of external platforms. Each adapter handles authentication, webhook management, and message formatting for its respective channel. Out of the box, OpenClaw ships with adapters for platforms such as:
- Slack, Microsoft Teams, and Discord
- Telegram, WhatsApp, and Signal
- Web chat, REST API, and email
- SMS via Twilio and Vonage
- And many more (50+ in total)
Skill Engine
The Skill Engine is responsible for discovering, loading, and executing skills. When the Agent Core determines that a user request requires a specific capability — for example, fetching the weather or querying a database — it delegates to the Skill Engine. Skills are self-contained modules with their own configuration, permissions, and dependency declarations.
Sandbox
The Sandbox provides an isolated execution environment for skills that run arbitrary code. It prevents a misbehaving skill from affecting the host system, other skills, or the Agent Core. By default, sandboxed skills run with restricted filesystem and network access.
Key concepts
Before you start building with OpenClaw, it helps to understand four foundational concepts.
Agents
An agent is a configured instance of the Agent Core. Think of it as your "virtual employee" or the "brain." Each agent has its own personality prompt, model selection, skill allowlist, and channel bindings. You can run multiple agents on a single OpenClaw installation — for example, one agent for customer support and another for internal DevOps tasks.
Skills
A skill is a discrete unit of functionality that an agent can invoke. Think of skills as the "hands and tools" given to the brain (e.g., a calculator or a web scraper). Skills range from simple (returning the current time) to complex (executing a multi-step API workflow). OpenClaw has thousands of community-maintained skills on ClawHub, and you can write your own using a simple Markdown format (SKILL.md with YAML frontmatter) or TypeScript for more complex integrations.
Channels
A channel is any communication platform where users interact with an agent. Think of this as the "office or desk" where your virtual employee works (like Slack, WeChat, or Telegram). Channels are pluggable: you enable only the ones you need, and each channel can be connected to one or more agents.
Models
A model refers to the large language model (LLM) that powers an agent's reasoning. This is the "intelligence engine" underlying the agent. OpenClaw is model-agnostic — it supports providers such as OpenAI, Anthropic, Google, Mistral, Ollama, and any OpenAI-compatible endpoint. You choose the model per agent, so one agent can use GPT-4o while another uses a locally hosted model.
Why OpenClaw?
There are several AI agent frameworks available today. Here is how OpenClaw compares to common alternatives.
| Feature | OpenClaw | Typical SaaS Bot Platforms | DIY Framework (e.g., LangChain) |
|---|---|---|---|
| Open source | Yes (MIT license) | No | Yes |
| Channel integrations | 50+ built-in | Varies (often 5-10) | You build them |
| Built-in skills | 5,700+ | Limited marketplace | You build them |
| Model agnostic | Yes | Usually locked to one provider | Yes |
| Self-hosted option | Yes | No | Yes |
| Managed hosting available | Yes | Yes | No |
| Sandboxed execution | Yes | Varies | No |
| Production ready | Yes | Yes | Requires significant work |
When to choose OpenClaw
- You want a batteries-included platform — hundreds of skills and dozens of channels work out of the box.
- You need to own your data — self-hosting means your conversations and API keys never leave your infrastructure.
- You support multiple channels — one agent definition, many frontends.
- You want model flexibility — swap providers without rewriting your agent logic.
When OpenClaw might not be the best fit
- You only need a simple chatbot widget with no integrations — a hosted SaaS tool may be faster to set up.
- You require a highly custom inference pipeline with fine-tuned models and no pre-built skills — a lower-level framework gives you more control.
Next steps
Now that you understand what OpenClaw is and how it works, head to the Requirements page to make sure your system is ready for installation.