OpenClaw

OpenClaw is a free and open-source personal AI assistant framework that runs on your own devices and communicates through the messaging platforms you already use. Created by Austrian developer Peter Steinberger, OpenClaw allows users to deploy an autonomous AI agent that connects to WhatsApp, Telegram, Discord, Slack, Signal, iMessage, and over 20 other messaging services, executing tasks via large language models (LLMs) such as Claude, GPT, and DeepSeek.

Unlike traditional chatbots, OpenClaw is designed as a local-first gateway - a persistent daemon that acts as a control plane for sessions, channels, tools, and events. Users interact with their agent through familiar messaging interfaces, while the agent runs autonomously in the background, capable of browsing the web, sending emails, controlling smart home devices, and executing complex multi-step workflows. The project is licensed under MIT and hosted on GitHub with over 381,000 stars as of mid-2026.

OpenClaw traces its roots to a project called Warelay, first published by Peter Steinberger in November 2025. The name "Clawd" was inspired by Anthropic's Claude chatbot, playing on the lobster-themed branding that became the project's hallmark.

Date	Event
Nov 2025	Warelay first published (predecessor)
Jan 2, 2026	Public release as Clawdbot
Jan 27, 2026	Rebranded to Moltbot (Anthropic trademark complaint)
Jan 30, 2026	Renamed to OpenClaw, final name
Feb 14, 2026	Steinberger joins OpenAI; OpenClaw Foundation established

Version Milestones

Version	Date	Key Changes
v0.x	Nov 2025	Early development, baseline agent functionality
v3.0	Nov 2025	Multi-model support, workspace system
v3.5	Jan 10, 2026	Voice integration, Playwright browser automation
v4.0	Feb 20, 2026	"The Agent OS", gateway daemon, 15+ platforms, cron

OpenClaw employs a hub-and-spoke architecture centered on a long-running daemon called the Gateway. This is fundamentally different from CLI-first agents; OpenClaw is designed as a persistent service that stays running and listening across multiple channels simultaneously.

The Gateway is a Node.js daemon that loads all messaging adapters directly into its own process and exposes a typed WebSocket API on ws://127.0.0.1:18789. Exactly one gateway runs per host, managing all agent sessions, channel connections, tool execution, and event routing. The wire protocol uses three frame types: req (client request), res (response), and event (server-side push).

Agent Runtime Loop

1. Input normalization - Channel adapters normalize incoming messages into a standard format.
2. Routing and session management - Messages are routed to the correct agent workspace and session.
3. LLM inference - The agent sends the prompt to the configured LLM provider with system prompt, history, and tool definitions.
4. Tool execution - When the LLM requests a tool call, the agent executes it and returns the result.
5. Response delivery - The final response is sent back through the originating messaging channel.

Companion Apps

• macOS app - Menu bar control, Voice Wake, push-to-talk overlay, WebChat, remote gateway management over SSH
• Windows Hub - Native desktop companion with tray status, chat, node mode, and local MCP mode
• iOS node - Pairs over WebSocket, voice trigger forwarding, Canvas surface
• Android node - Connect/Chat/Voice tabs, Canvas, Camera, Screen capture, device command families

• Multi-Channel Inbox - Over 20 messaging platforms as first-class interfaces: WhatsApp, Telegram, Slack, Discord, Google Chat, Signal, iMessage, IRC, Teams, Matrix, and more.
• Local-First Gateway - All data processing, configuration, and history are stored locally. No cloud dependency beyond LLM API calls.
• Skills System - Extend capabilities through SKILL.md files with YAML frontmatter. Skills can be bundled, managed from ClawHub, or workspace-created.
• Voice and Talk Mode - Voice Wake on macOS/iOS with configurable wake words, Talk Mode on Android, ElevenLabs TTS, Whisper STT.
• Live Canvas - Agent-driven visual workspace with interactive UI components (A2UI), data visualizations, and real-time user interaction.
• Cron Jobs and Automation - Built-in scheduler for recurring tasks, webhook support, and Gmail Pub/Sub integration.
• Agent Protocol (ACP) - Standardized protocol for client-agent communication with headless CLI via acpx.
• Multi-Agent Routing - Route channels, accounts, and peers to isolated agents through workspaces with separate configs and skills.

Prerequisites: Node.js 24 (or Node.js 22.19+), npm/pnpm/bun, and an LLM API key.

Standard Install:

npm install -g openclaw@latest
pnpm add -g openclaw@latest
openclaw onboard --install-daemon

The onboard command guides through gateway configuration, workspace creation, channel setup, LLM provider config, and skill installation step by step.

Getting Started

openclaw gateway start          # Start the gateway daemon
openclaw gateway status         # Check gateway health
openclaw agent --message "Hello"  # Send your first message

Chat Commands

Control the agent through built-in commands from any messaging channel:

Command	Description
/status	Check agent status and health
/new	Start a fresh conversation session
/reset	Reset the current session
/compact	Compress conversation history
/think level	Set thinking effort (low/medium/high)
/restart	Restart the agent session

OpenClaw's most distinctive feature is its universal inbox for AI agent interactions. The Gateway loads messaging adapters natively, integrating each platform directly into the daemon process rather than proxying through external services.

Category	Platforms
Consumer Messaging	WhatsApp, Telegram, Signal, iMessage, WeChat, QQ
Team Collaboration	Slack, Discord, Microsoft Teams, Google Chat, Mattermost, Nextcloud Talk
Developer/Open	IRC, Matrix, Nostr, Zalo, LINE
Other	Tlon, Twitch, Synology Chat, WebChat, Feishu, macOS notifications

The skills system is the primary extension mechanism. Each skill is a directory containing a SKILL.md file with YAML frontmatter (metadata, triggers, dependencies) and a markdown body with instructions, tool definitions, and examples.

Type	Location	Priority
Bundled	openclaw/dist/skills/	Lowest
Managed	~/.openclaw/skills/	Medium
Workspace	~/.openclaw/workspace/skills/	Highest

ClawHub (clawhub.ai) is the community skill marketplace. Install skills with: openclaw skill install NAME.

OpenClaw is configured through a JSON file at ~/.openclaw/openclaw.json. Key configuration sections include: agent (model selection, defaults, sandbox mode), gateway (port, host, daemon settings), channels (per-channel config), agents (multi-agent routing), security (DM policies, sandboxing, allowlists), cron (scheduled jobs), and webhooks.

OpenClaw also reads injected prompt files from the workspace root: AGENTS.md (system-level instructions), SOUL.md (agent persona and character), and TOOLS.md (custom tool definitions).

OpenClaw operates on a principle of default-deny for external access. Unknown senders receive a pairing code; the bot does not process their message until approved. The agents.defaults.sandbox.mode: "non-main" setting runs non-main sessions inside Docker, SSH, or OpenShell sandboxes.

Several CVEs have been documented for OpenClaw (CVE-2026-25253, CVE-2026-25891, CVE-2026-26102) related to prompt injection and input sanitization. The project's maintainers emphasize that OpenClaw requires technical competence to operate safely. Cisco's AI security team found that ClawHub's skill repository lacked adequate vetting. In March 2026, Chinese authorities restricted OpenClaw on state-run enterprise computers due to security concerns.

Aspect	OpenClaw	Hermes Agent
Architecture	Hub-and-spoke Gateway daemon	CLI-first with optional gateway
Design philosophy	Persistent always-on assistant	On-demand task execution
Messaging	Native multi-channel (20+ platforms)	Gateway plugin model (12+ platforms)
Skills/Learning	Static SKILL.md files	Self-improvement loop, creates skills from experience
Memory	Conversation-based, file-backed	SQLite with multiple backend plugins
Ecosystem	Larger third-party skill ecosystem	Stronger Nous Research model integration
GitHub stars	381K	210K

• Personal AI Assistant - Email management, calendar scheduling, web research, document drafting through WhatsApp, Telegram, or iMessage.
• Business Process Automation - Lead generation, prospect research, website auditing, CRM integration, and follow-up email sequences.
• Developer Productivity - Always-available coding assistant that executes shell commands, browses documentation, manages git repos, and coordinates tools.
• Automated Content Pipelines - Daily news digests, social media posting, data collection, and recurring report generation via cron scheduling.
• Smart Home Control - Voice-activated controller for lights, thermostats, and IoT devices through messaging interfaces.