10 mins read
What is OpenClaw (previously Moltbot and Clawdbot)?
OpenClaw is an open-source, self-hosted AI agent framework that transforms large language models into proactive personal assistants capable of automating real-world tasks directly on your devices. Launched in late 2025 by PSPDFKit founder Peter Steinberger, it quickly became one of the fastest-growing open-source projects, amassing over 196,000 GitHub stars by mid-February 2026 amid viral developer excitement and multiple rebrands due to trademark issues. Originally dubbed Clawdbot, it briefly shifted to Moltbot before settling on OpenClaw in January 2026, earning nicknames like "Molty" or "lobster" from its community for its resilient, tool-wielding design reminiscent of a crustacean's claws.
Unlike traditional chatbots that merely respond to queries, OpenClaw acts as a "24/7 JARVIS-like" agent, bridging AI models with your local environment to handle tasks autonomously—such as managing emails, controlling smart home devices, scraping web data, or even negotiating purchases. It runs as a lightweight Node.js service on macOS, Windows, or Linux, emphasizing privacy by keeping data local and allowing users to bring their own API keys from providers like OpenAI, Anthropic, or local models via Ollama. By February 2026, Steinberger announced his move to OpenAI to advance personal agents, while transferring OpenClaw's stewardship to an independent foundation to ensure its open-source future.
What sets OpenClaw apart is its agentic architecture: it maintains persistent memory across sessions using local Markdown files, self-improves by generating new "AgentSkills" (over 100 pre-built bundles for tasks like shell execution or web automation), and integrates with 50+ tools including WhatsApp, Discord, Gmail, GitHub, Spotify, and Philips Hue. Users interact via familiar chat apps, granting it sandboxed or full system access to read/write files, run scripts, or control browsers—enabling scenarios like automated debugging, health tracking from wearables, or even building full apps while you're away. Recent February 2026 updates have supercharged it further: version 2026.2.19 added Apple Watch support for wrist-based triggers, while 2026.2.21 introduced massive security hardening, Gemini 3.1 integration, Discord voice streaming, and over 100 fixes to address vulnerabilities like CVE-2026-25253 (a one-click remote code execution flaw patched earlier). A new partnership with VirusTotal enhances skill scanning for malware, responding to incidents where 341 malicious skills infiltrated its marketplace in late January.
OpenClaw remains completely free and open-source under the MIT license, with costs limited to your chosen AI model's API usage—making it accessible for hobbyists and enterprises alike. For those seeking managed deployments without the setup hassle, services like Clawnify.com build directly on OpenClaw, offering fully hosted AI "employees" with dedicated infrastructure, real-time monitoring, and app-building tools starting in just 5 minutes—ideal for scaling agentic workflows in teams via Slack or WhatsApp. Community-driven growth has led to innovative use cases, from India's first multi-city OpenClaw Buildathon generating revenue-building agents to users cautioning about over-eager automations (like accidental inbox deletions). As of February 23, 2026, OpenClaw's blend of autonomy, extensibility, and privacy positions it as a foundational tool for the agentic AI era, though users should prioritize security best practices amid its rapid evolution.
How OpenClaw works
OpenClaw operates as a lightweight, persistent Node.js gateway service that bridges large language models (LLMs) with your local environment and messaging apps, enabling autonomous task execution while prioritizing privacy and extensibility. Unlike passive chatbots, it functions as an "AI operating system" or proactive agent runtime, running 24/7 on macOS, Windows, Linux, or VPS setups (e.g., via DigitalOcean, Hetzner, or Raspberry Pi). This setup allows it to maintain context across sessions, self-improve through code generation, and perform real-world actions like file management, browser automation, or API calls—all without sending data to external servers unless explicitly configured.
At its core, OpenClaw uses an agentic loop: incoming messages from integrated channels (e.g., WhatsApp, Telegram, Discord, Slack, Signal, iMessage, or even custom webhooks) are routed through a central Gateway, which interprets user intent via your chosen LLM (model-agnostic, supporting OpenAI, Anthropic's Claude, Gemini, local models like Llama via Ollama, or cost-efficient options like MiniMax 2.1). The LLM then decides on actions, leveraging tools and skills to execute tasks in a sandboxed or full-access mode. A key differentiator is the "heartbeat" mechanism—a scheduled cron-like loop where the agent proactively checks for tasks, scans environments (e.g., emails, calendars, or external APIs), and initiates actions without user prompts, such as sending daily summaries or monitoring health metrics from wearables.
Memory management is file-based and persistent: conversations, preferences, and states are stored in local Markdown files or JSONL transcripts within a workspace directory, allowing for hybrid semantic/keyword retrieval (e.g., via grep or advanced indexing). This enables long-term context retention, cross-session continuity, and easy manual edits—far beyond volatile chat histories. For extensibility, OpenClaw supports over 100 community-contributed "AgentSkills"—modular, AI-generatable bundles for specific workflows like shell commands, web scraping via Playwright (with built-in vision for screen-to-text conversion), or integrations with Gmail, GitHub, Spotify, Philips Hue, Todoist, or even custom CLIs. Skills can be self-created: prompt the agent to write and install new ones, complete with guardrails and tool calls.
Multi-agent collaboration is a February 2026 highlight: the system can spawn sub-agents for parallel tasks (e.g., one for research, another for execution), with handoffs via shared memory or channels, mimicking team workflows. For instance, a primary agent might delegate browser control to a sub-agent while another handles data processing, all coordinated through the Gateway's Message Control Plane (MCP) for routing and isolation. Sessions can run sequentially, concurrently, or in batched "collect" modes, with isolation keys encoding agent scopes to prevent context pollution.
Security remains a focus amid rapid growth: data stays local by default, with user-controlled permissions (sandbox vs. full access). As of February 2026, integrations with VirusTotal scan skills for malware, addressing incidents like the late-January infiltration of 341 malicious extensions in the marketplace. Users are advised to use containerized setups (e.g., via Docker or E2B sandboxes) for high-risk tasks, and recent patches (e.g., v2026.2.21) fixed vulnerabilities like CVE-2026-25253 for remote code execution. Potential risks include over-automation (e.g., accidental file deletions), prompt injections, or resource overconsumption from unchecked loops—mitigated by audit trails, human-in-the-loop checkpoints, and token budgeting.
For a high-level overview of core components:
Component | Description | Key Benefits | Recent Updates (Feb 2026) |
|---|---|---|---|
Gateway | Central router for messages, sessions, and tasks; handles channel adapters for apps like WhatsApp/Slack. | Unifies interfaces; enables proactive heartbeats for 24/7 operation. | Enhanced multi-agent routing with isolation keys to prevent cross-contamination. |
Brain/Agent Runtime | LLM-powered decision engine; selects tools/skills based on intent. | Model-agnostic (e.g., Claude 4.6 for coding, MiniMax for low-latency); supports self-hacking for prompt edits. | Integration with Gemini 3.1 and voice streaming in Discord. |
Memory System | File-backed Markdown/JSONL storage with retrieval layers. | Persistent, editable context; supports long-term "second brain" building. | Hybrid search improvements for faster recall in multi-agent setups. |
Skills & Tools | Modular extensions (e.g., Playwright for browser, shell execution). | AI-generated; over 100 bundles for web automation, APIs, smart home. | VirusTotal scanning; new skills for Apple Watch triggers and JMAP email search. |
Heartbeat Loop | Scheduled proactive checks for tasks/environment monitoring. | Shifts from reactive to autonomous (e.g., daily reports, reminders). | Optimized for background efficiency, reducing token waste by 90% in some workflows. |
In practice, a workflow might start with a WhatsApp message like "Buy a flashlight on Amazon": the Gateway routes it to the LLM, which invokes Playwright to navigate, search, and complete the purchase—transcribing audio if needed via external APIs—all while updating memory for future preferences. For teams or scaled use, services like Clawnify.com extend this by providing managed, multi-agent deployments on dedicated infrastructure—complete with real-time monitoring, Slack/WhatsApp integration, and one-click app building—eliminating local setup hassles while maintaining OpenClaw's core autonomy. As of February 23, 2026, OpenClaw's architecture continues evolving toward a "multi-agent economy," with community forks like NanoClaw emphasizing auditable, containerized simplicity for security-conscious users.
Start using OpenClaw in an isolated computer
Make you first hire
How to set up OpenClaw
Setting up OpenClaw is straightforward thanks to its streamlined CLI installer and onboarding wizard, making it accessible for beginners while offering advanced options for production environments. As of February 2026, the process emphasizes security from the start, with built-in prompts for sandbox mode and VirusTotal skill scanning to mitigate risks like the January marketplace infiltration. It runs on macOS, Windows (via WSL2 for best results), Linux, or cloud VPS, requiring Node.js 22+—which the installer handles automatically. The core setup takes under 10 minutes for local use, but for 24/7 operation, a dedicated VPS (e.g., Hetzner at €4/month or Tencent Cloud Lighthouse at 6¥/month) is recommended over consumer hardware like Mac Minis to avoid power and firewall issues.
Before diving in, consider your goals: for personal testing, a local install suffices; for always-on agents handling tasks like lead generation or automation, opt for cloud deployment to enable proactive heartbeats without keeping your machine running. Security is paramount—always start in sandbox mode, back up configs, and set up alerts for failed logins, disk usage, and unauthorized changes, as brute-force attempts have spiked in early 2026. If self-hosting feels overwhelming, services like Clawnify.com offer a fully managed alternative: deploy an AI "employee" in 5 minutes via Slack integration, with dedicated infrastructure, one-click OAuth for tools (e.g., Gmail, HubSpot), real-time screen monitoring, and built-in app building—no terminal required. This bypasses common pitfalls like dependency debugging or VPS firewall tweaks, ideal for teams scaling agentic workflows.
Here's a step-by-step guide to core setup, with variations for platforms and advanced options:
Check Prerequisites: Ensure Node.js 22+ is installed (node --version); if not, the installer adds it. For Windows, enable WSL2 via Settings > Apps > Optional Features > Windows Subsystem for Linux. Have API keys ready for LLMs (e.g., Anthropic, OpenAI, or local via Ollama for free setups like GLM 4.7 Flash).
Run the Installer: Open a terminal/PowerShell and execute the one-liner. This fetches the latest version (v2026.2.21+ as of Feb 23), installs globally via npm, and sets up the daemon service for background running.
macOS/Linux:
curl -fsSL https://openclaw.ai/install.sh | bashWindows (PowerShell):
iwr -useb https://openclaw.ai/install.ps1 | iexAlternative (manual):
npm install -g openclaw@latest
Launch Onboarding Wizard: Run
openclaw onboard --install-daemon(or justopenclawif prompted). Select "QuickStart" mode: configure LLM provider (e.g., Claude 4.6 for reasoning, Codex 5.3 for coding), paste API keys, and link chat channels like Telegram (via @BotFather token) or WhatsApp. Enable sandbox for security; add skills from the registry (e.g., browser control via Playwright).Test and Customize: Send a test message via your chosen app (e.g., "Hello" in Telegram). Edit USER.md in the workspace (~/.openclaw/) for personality/preferences. For local models, integrate Ollama first: install via their site, pull GLM 4.7 Flash (~25GB), then select it in the wizard.
Advanced/Cloud Setup: For VPS (e.g., Ubuntu on Hetzner/Hostinger), SSH in, update packages (
sudo apt update && sudo apt upgrade), install Node.js if needed, then follow steps 2-4. Use Docker for isolation: clone repo, run setup script. DigitalOcean's 1-Click Deploy (starting at $24/month) pre-installs v2026.1.24+ with hardened security. For managed ease, use Clawnify: select a role/template, connect via OAuth, and deploy—no code needed.
Platform/Method | Key Command/Steps | Best For | Common Issues & Fixes (Feb 2026) |
|---|---|---|---|
Local (macOS/Linux) | curl install > onboard wizard | Quick testing, personal use | macOS permissions: Allow in System Settings; slow on M1-M5: Use Rosetta for Node if needed. |
Windows (WSL2) | PowerShell iex > onboard | Developers avoiding native Windows quirks | Telegram pairing fails: Restart WSL, check firewall; use Ubuntu distro for stability. |
VPS (e.g., Hetzner Ubuntu) | SSH > apt update > curl install > onboard | 24/7 operation, low cost (~€4/mo) | Firewall blocks: ufw allow 22,3000; disk full: Monitor with alerts, expand storage. |
Docker/Containerized | git clone > run setup script | Isolated, secure environments | Container resets: Use volumes for persistence; CVE fixes in v2026.2.21. |
Managed (Clawnify) | Pick role > OAuth connect > Slack tasks | Teams, no-dev setups (5 min) | None—handled; custom needs: Use Expert View for OpenClaw access. |
Cloud 1-Click (DigitalOcean) | Search Marketplace > Deploy Droplet > Configure | Production-ready, hardened | Version lag: Update post-deploy; costs start $24/mo. |
Troubleshooting highlights from Feb 2026: Telegram pairing issues (restart bot/service), memory forgetting (compact Markdown files manually if auto-flush bugs), high costs (switch to Kimi at $0.60/M tokens with caching). For enterprise, follow hardened checklists: dedicated machines, SSH alerts, daily audits. Once set up, OpenClaw's wizard makes expansions easy—prompt it to install skills or optimize configs, turning setup into an agent-assisted process.
How to use OpenClaw
Once set up, OpenClaw transforms into a versatile, always-on AI companion that handles tasks via natural language in your preferred messaging apps, leveraging its persistent memory and extensible skills to automate workflows with minimal oversight. Interaction is conversation-first: message it in WhatsApp, Telegram, Discord, Slack, Signal, or iMessage (including group chats), and it responds proactively, drawing from stored context in local Markdown files like USER.md for preferences or SOUL.md for core instructions. As of February 23, 2026, recent enhancements include oAuth for Anthropic subscriptions (run openclaw models auth add to integrate Claude 4.6 without API costs), voice streaming in Discord, and multi-agent orchestration where a primary agent delegates to sub-agents for parallel execution—boosting throughput for complex tasks like coding or research. For teams, managed platforms like Clawnify extend this by deploying "AI employees" in Slack or WhatsApp, with one-click OAuth for tools like Gmail or HubSpot, real-time screen monitoring, and app-building capabilities—all without manual hosting.
Start simple: Send "Hello" in your connected app to test; then escalate to prompts like "Summarize my emails from today" or "Build a dashboard for my leads." OpenClaw excels in proactive mode via its heartbeat loop (configured in HEARTBEAT.md), running scheduled checks every few minutes to monitor environments, trigger automations, or send updates—e.g., daily briefings aggregating Slack, email, calendar, and news. Edit system files directly for customization: tweak IDENTITY.md for persona (e.g., "Jarvis" or "Claudia"), or add cron jobs for recurring tasks like SEO reports or competitor monitoring. For security, enable human-in-the-loop via confirmations for sensitive actions, and use VirusTotal-integrated scanning for new skills to avoid malware risks seen in January's marketplace incident.
OpenClaw's 50+ integrations span chat platforms, LLMs (Claude, GPT, Gemini, local like MiniMax 2.1), productivity tools (Gmail, GitHub, Obsidian, Todoist, Trello), media (Spotify, Sonos), smart home (Philips Hue, Home Assistant, Winix air purifiers), and more—enabling seamless bridging between AI and real-world actions. Browser control via Playwright allows web automation like form filling or scraping, while shell access handles file ops, scripts, or even PC shutdown. Recent additions include Apple Watch triggers for wrist-based commands and JMAP for advanced email search.
OpenClaw Use Cases
OpenClaw shines in personal, developer, and business scenarios, shifting from reactive chats to autonomous operations that save hours weekly. Key categories include:
Personal Productivity: Manage inboxes (clear 10k+ emails, unsubscribe spam), calendars (traffic-based reminders, flight check-ins), health (pull WHOOP data, track biomarkers), and daily routines (grocery orders, custom meditations with TTS/audio).
Developer Workflows: Automate debugging, code reviews, test fixes (e.g., via Telegram progress updates), PR creation from Sentry webhooks, overnight coding with sub-agents, or full app builds/deployments.
Content & Marketing: Repurpose blogs into threads/posts, monitor audiences/competitors, generate visuals/scripts, or syndicate content across platforms.
Research & Monitoring: Continuous crawls of Reddit/X/HN, SEO analysis, document Q&A, or private CRM for relationship tracking/follow-ups.
Finance & Admin: Spending audits, insurance disputes (draft rebuttals), receipt parsing, or commercial real estate zoning analysis.
Smart Home & Media: Control lights/air purifiers, generate videos (with watermark removal), or phone calls with accents.
Community-driven examples as of February 2026 highlight real impact: One user negotiated $4,200 off a car via email overnight; another built a $4k/week business with lead scraping/DM routing; families save hours on meal planning in Notion; devs run autonomous loops for error resolution. In teams, Clawnify automates sales leads, support queries, or research, with AI "employees" building/deploying apps like dashboards—scaling workflows profitably without code.
OpenClaw for AgentSkills
Expand capabilities with 100+ pre-built AgentSkills—modular bundles for workflows like shell execution, web scraping, or custom CLIs—or prompt OpenClaw to generate new ones (e.g., "Create a skill for querying flights"). Install via registry search and terminal commands, or let the agent self-install with guardrails. Recent skills include multi-agent handoffs, voice-to-text, and RSS parsing; community shares 324+ use cases on sites like Clawledge.
Category | Example Skill | Prompt Example | Benefits & Tips (Feb 2026) |
|---|---|---|---|
Automation | Shell Executor | "Run this script to backup files" | Enables cron jobs; optimize with token caching to cut costs 90%. |
Web | Playwright Browser | "Scrape data from site X" | Vision-enabled for screenshots; use sub-agents for parallel tabs. |
Productivity | Gmail/Todoist | "Clear my inbox" | OAuth integration; edit memory to prevent over-automation errors. |
Media | Spotify/TTS | "Create meditation audio" | Generate content; proxy for subscriptions to avoid API fees. |
Custom | AI-Generated | "Build a Stumbleupon clone" | Self-hacks prompts; scan with VirusTotal before use. |
For advanced users, explore forks like PicoClaw on Arduino for hardware integrations or GodMode for concierge-style ops. If self-hosting limits scale, Clawnify's managed agents provide enterprise-grade visibility and ease, turning OpenClaw into team-ready AI without the grind. As community feedback notes, real value emerges over time—start small, iterate, and watch it evolve into your "second brain."
