What Is OpenClaw
OpenClaw is an open-source personal AI assistant that runs on your device and responds in the messengers you already use. The project has gathered over 380,000 GitHub stars and became one of the fastest-growing open-source repositories in history. Below is what this tool is, how it works, and who it is for.
What the project is
OpenClaw (formerly Moltbot and Clawdbot) is a self-hosted agent created by Peter Steinberger (founder of PSPDFKit). The project mascot is the space lobster Molty 🦞. The idea is simple: instead of yet another chat window, the assistant lives where you already spend time - WhatsApp, Telegram, Slack, Discord, Signal, iMessage, and dozens of other channels.
OpenClaw does not sell a model. You connect your own LLM - OpenAI, Anthropic, Google, local models via Ollama, or any compatible provider. A gateway on your machine routes messages, runs skills, and performs real actions: shell commands, file operations, browser automation, API calls, calendar, smart home.
The project is written in TypeScript, distributed as open source, and actively developed by the community (370+ contributors, releases roughly every two days).
How the architecture works
At the center is the Gateway - a long-running Node.js service that acts as the control plane. The product is the assistant, not the gateway: the Gateway connects communication channels to the agent and workspace.
Main components:
- Gateway - message routing, device pairing, session management
- Agent - assistant core: reasoning, tool calls, memory
- Skills - extensible capabilities; the agent can write code and add new skills for your tasks
- Channels - adapters for messengers and chat platforms
- CLI (
openclaw) - installation, onboarding, diagnostics, daemon management
Configuration is stored in a single file ~/.openclaw/openclaw.json (JSON5 format). This simplifies backup and migration between machines.
macOS, Linux, and Windows are supported. On Windows there is a native Windows Hub app for setup, tray status, chat, and local MCP mode.
Key capabilities
Messengers instead of a separate UI
OpenClaw connects to 50+ channels: WhatsApp, Telegram, Slack, Discord, Google Chat, Signal, iMessage, Microsoft Teams, Matrix, IRC, Feishu, LINE, Mattermost, WebChat, and more. You message the assistant from a familiar app - it replies and runs tasks on your machine.
Local-first and data control
Data and logs stay on your hardware. LLM requests go only to the provider you choose - you can use local models and avoid sending work documents to a cloud chat.
Autonomous actions
The agent is not limited to text replies. It can:
- run terminal commands
- read and modify files in the workspace
- automate the browser
- call external APIs
- schedule tasks and reminders
- manage integrations (calendar, smart home, CRM)
Voice, Canvas, and mobile clients
On macOS, iOS, and Android, voice input and spoken replies are supported. Canvas is a live visual layer controlled by the assistant (useful for dashboards, forms, previews).
Skills system and self-improvement
Skills are modular extensions of agent behavior. The community publishes ready-made skills; the assistant can write a new skill for a recurring task and keep it across sessions. This sets OpenClaw apart from one-off chatbots.
Model-agnostic
A current flagship from a trusted provider is recommended, but the architecture is not tied to one model. The issue tracker actively discusses setups with Qwen 3.5-Coder, Hermes, Gemma 4, GLM-5, and other local models.
Quick start
Requirements: Node 24 (recommended) or Node 22.19+.
npm install -g openclaw@latest
openclaw onboard --install-daemon
The onboard command walks through setup step by step: gateway, workspace, channels, skills, and pairing. The --install-daemon flag installs a background service (launchd on macOS, systemd on Linux) so the assistant runs 24/7.
Check status:
openclaw gateway status
For debugging, you can run the gateway in the foreground:
openclaw gateway --port 18789 --verbose
Documentation: docs.openclaw.ai, website: openclaw.ai.
Who OpenClaw is for
| Scenario | Why OpenClaw |
|---|---|
| Personal automation | Assistant in Telegram/WhatsApp runs tasks on your PC |
| Privacy-first workflow | Self-hosting, model choice, workspace control |
| Power user / developer | Shell, API, skills, extensible architecture |
| Always-on helper | Daemon + messengers = access from your phone anytime |
| Local LLMs | Ollama and open weights without vendor lock-in |
It fits less well where you need enterprise SLA, centralized IAM, and ready managed SaaS without DevOps. OpenClaw is a tool for a single user (or a small self-hosted setup), not an out-of-the-box corporate platform.
Pros and limitations
Strengths:
- explosive growth of the ecosystem and community skills
- unique UX through messengers
- open source and portability across OSes
- flexible LLM providers
- active releases and mature CLI setup
Things to watch:
- security depends on your configuration: an agent with shell access needs deliberate workspace limits
- the project is young - APIs and best practices change quickly
- channel setup (especially WhatsApp, iMessage) can take time
- complex tasks still need a strong model - local LLMs are catching up but are not everywhere at the level of GPT-5.6 Sol or Claude Fable 5
OpenClaw in the 2026 context
The AI agent market has shifted from demo chats to work tools: ChatGPT Work, Claude Cowork, Gemini Spark - products from major labs. OpenClaw occupies a different niche - a personal, local, messenger-centric assistant without a platform subscription.
The rise of local-first models (Qwen 3.6, DeepSeek V4, Gemma 4, GLM-5) made self-hosted agents more practical: quality on real tasks is closer to closed APIs, and OpenClaw already has infrastructure for that wave.
If you want a 24/7 Jarvis on your own hardware with access from WhatsApp at 2 AM - OpenClaw is today one of the most mature open-source options in this category.
Frequently asked questions
Is OpenClaw a chatbot or a full agent?
A full agent. Beyond chat replies, OpenClaw performs actions on your machine: terminal, files, browser, API, skills from the ecosystem. The gateway only links messengers to the agent - the product is the assistant itself.
Do I need a paid OpenClaw subscription?
No. OpenClaw itself is free and open source. You only pay for an LLM provider if you use a cloud API (OpenAI, Anthropic, Google). With local models via Ollama, costs boil down to power and hardware.
Which messengers are supported?
More than 50 channels: WhatsApp, Telegram, Slack, Discord, Signal, iMessage, Google Chat, Microsoft Teams, Matrix, IRC, Feishu, LINE, Mattermost, WebChat, and more. The full list is in the documentation.
Can I use local models without the cloud?
Yes. OpenClaw is model-agnostic: Ollama and other local runtimes connect. In 2026 many run Qwen 3.5-Coder, Hermes, Gemma 4, or GLM-5 for coding and routine automation without sending data to the cloud.
Is it safe to give the agent shell access?
An agent with shell access is powerful but risky. Recommendations: limit the workspace to a separate directory, do not run as root, review skills before installing, use pairing for channels, update OpenClaw regularly. For production-critical environments, isolate the gateway in a VM or container and explicitly whitelist allowed tools.