OpenClaw is the most popular open-source AI assistant platform on GitHub, with over 335K stars and a growing community of self-hosters who refuse to hand their data to Big Tech. If you have heard about it but never quite understood what it does or how it works, this guide covers every angle.
A Brief History of OpenClaw
The project started in late 2023 under the name Clawdbot, a hobby experiment by Peter Steinberger to connect an LLM to a Telegram chat. Within months it gained traction, was renamed to Moltbot, and then settled on the name OpenClaw in mid-2024 after the community voted on GitHub Discussions.
By January 2025 it had crossed 100K stars. Today, in March 2026, it sits at 335K+ stars, making it the fastest-growing open-source AI project after llama.cpp. Contributions come from over 1,200 developers, and the release cadence averages two minor versions per month.
What OpenClaw Actually Does
At its core, OpenClaw acts as a self-hosted bridge between your favourite messaging apps and any large language model (LLM). You install it on a server, connect one or more chat channels, pick an AI model, and you have your own private AI assistant that remembers context, runs skills (plugins), and stays entirely under your control.
Unlike ChatGPT or Claude.ai where your conversations live on someone else's servers, OpenClaw stores everything on your own infrastructure. You choose the database, you own the encryption keys, and you decide which AI provider processes your prompts.
The Gateway Architecture
OpenClaw uses a Gateway pattern that separates concerns cleanly. The Gateway layer handles authentication, rate limiting, and channel routing. Behind it, the Orchestrator decides which model to call, how to build the prompt, and whether to invoke any skills. Finally, the Memory layer handles context persistence.
- Gateway — Receives messages from WhatsApp, Telegram, Discord, Slack, or the web UI. Normalises them into a common format.
- Orchestrator — Routes the normalised message to the right LLM, appends relevant memory, and applies any active skills.
- Memory — Stores conversation history, user preferences, and long-term facts. The default engine uses SQLite, but PostgreSQL and Redis adapters exist.
- Skills — Plugin system with 191+ community skills for tasks like web search, calendar integration, home automation, and code execution.
Supported AI Models in March 2026
OpenClaw is model-agnostic. You bring your own API key, and the Orchestrator routes to whichever provider you configure. The currently supported models include:
| Provider | Models | Approx. Cost per Message |
|---|---|---|
| Anthropic | Claude Opus 4.6, Sonnet 4.6, Haiku 4.5 | $0.05 / $0.01 / $0.001 |
| OpenAI | GPT-5, GPT-5 mini | $0.03 / $0.002 |
| Gemini 3 Pro, Gemini 3 Flash | $0.02 / $0.001 | |
| DeepSeek | DeepSeek V3 | $0.005 |
| Mistral | Mistral Large | $0.008 |
| Meta | Llama 4 (via Ollama, local) | Free (your hardware) |
| Alibaba | Qwen 3 (via Ollama, local) | Free (your hardware) |
Running a local model like Llama 4 or Qwen 3 through Ollama eliminates API costs entirely, though you need a machine with at least 16 GB of RAM for the smaller quantised variants.
Messaging Channels
One of OpenClaw's strongest selling points is multi-channel support. A single instance can serve all of these simultaneously:
- WhatsApp (via the official Cloud API or third-party bridges)
- Telegram (native bot API)
- Discord (bot integration)
- Slack (app integration)
- Web UI (built-in React frontend)
- Matrix / Element
- Signal (community bridge, experimental)
Memory persists across channels. If a user talks to the assistant on Telegram in the morning and switches to WhatsApp in the afternoon, the conversation context carries over seamlessly.
The Skills Ecosystem
Skills are modular plugins that extend what OpenClaw can do. The community registry lists 191 skills at the time of writing. Popular categories include:
- Web search and scraping
- Calendar and scheduling (Google Calendar, Outlook)
- Smart home control (Home Assistant, Tuya)
- Code interpreter (sandboxed Python execution)
- Image generation (Stable Diffusion, DALL-E 3)
- RAG document retrieval (upload PDFs, query them)
The Supermemory System
OpenClaw's 6-layer memory architecture, branded Supermemory, achieves an 85.9% recall rate in independent benchmarks. That compares to 58.3% for the classic single-context approach most chatbots use. The six layers are:
- Short-term buffer (current conversation window)
- Session summary (compressed version of recent chats)
- User profile (extracted preferences and facts)
- Episodic memory (time-stamped notable events)
- Semantic index (vector embeddings for similarity search)
- Knowledge base (RAG over uploaded documents)
This layered approach allows OpenClaw to recall something a user mentioned weeks ago without stuffing the entire history into the prompt context. It reduces token usage and keeps API costs lower.
Security Concerns You Should Know About
OpenClaw's popularity has created a security problem. A scan performed in January 2026 found 42,665 OpenClaw instances exposed on the public internet. Of those, 93.4% had no authentication enabled, meaning anyone could read conversation history, change settings, or inject malicious prompts.
The ClawHavoc vulnerability (CVE-2026-25253) disclosed in February allowed remote code execution through a crafted skill payload. It affected all versions before v2026.2.5. The OpenClaw team patched it within 48 hours, but thousands of instances remain unpatched.
If you run OpenClaw, read our security hardening guide and run through the 12-point security checklist.
What Does It Actually Cost to Run?
The software itself is free and open source. Your costs come from two areas: hosting and AI API usage. A basic VPS starts at $5/month. AI costs depend on which model you choose and how many messages you send.
A light user sending 30 messages a day with GPT-5 mini pays roughly $5 VPS + $2 API = $7/month. A heavy user on Claude Opus 4.6 with 200+ daily messages might spend $24 VPS + $300 API = $324/month. Most people land somewhere in the $15-50/month range.
For a detailed cost breakdown, see our complete cost analysis.
Getting Started with OpenClaw
The fastest path from zero to a working assistant takes about 45 minutes. You need a VPS with at least 2 vCPU and 4 GB RAM, an AI API key, and a messaging channel token.
- Provision a VPS from Hostinger or DigitalOcean
- SSH in and run the one-line installer:
curl -fsSL https://get.openclaw.dev | bash - Follow the interactive setup to pick a model and connect a channel
- Harden your instance using our security checklist
For the full walkthrough with screenshots and troubleshooting tips, read our installation guide or the step-by-step VPS tutorial.
Should You Use OpenClaw or a Hosted Service?
OpenClaw is perfect if you value data ownership, want unlimited customisation, and are comfortable managing a server. If you prefer a hands-off experience, managed services like OpenClawPro handle the installation, security, updates, and monitoring for you.
The DIY approach saves money if you already have sysadmin skills. The managed approach saves time and reduces risk. There is no wrong answer, only trade-offs.
Looking Ahead
The OpenClaw roadmap for Q2 2026 includes voice channel support (phone calls via Twilio), a visual workflow builder for skill chains, and a built-in fine-tuning pipeline for local models. The project shows no signs of slowing down.
Whether you are a developer, a business owner, or just someone who wants a private AI assistant that does not sell your data, OpenClaw is worth your attention. Read our What is OpenClaw page for a condensed overview, or jump straight to the pricing page to see how OpenClawPro can get you running in a day.