What Is OpenClaw?
OpenClaw is an open-source personal AI assistant you run on your own machines. It is TypeScript, MIT-licensed, and has ~359k GitHub stars as of April 2026, making it one of the fastest-growing self-hosted AI projects on GitHub.
OpenClaw keeps conversations, data, and model weights on hardware you control. You pick the inference backend (local GPU with Ollama, a NIM service, vLLM, or a cloud API like Claude or OpenAI) and connect it to whichever messaging platforms your team already uses.
Core Architecture
OpenClaw is built as a modular gateway with three layers:
- Gateway. Receives messages from connected platforms and routes them to the inference engine. This is the minimum you need to run.
- Agent runtime. Executes tasks: browser automation, file operations, API calls, and shell commands. Runs in an isolated per-session container.
- ClawHub. A community registry of 5,700+ skills (plugins) that extend the agent runtime with pre-built capabilities.
All three layers communicate over a local message bus. The dashboard at :18789 is a web UI for managing connections, reviewing session history, and installing ClawHub skills.
Messaging Platform Integrations
OpenClaw connects to messaging platforms as a first-class feature. You talk to it through the same apps you already use, with no new UI to adopt.
| Platform | Support Level |
|---|---|
| Full (QR-linked session) | |
| Telegram | Full (bot token) |
| Slack | Full (bot + app token) |
| Discord | Full (bot token) |
| Signal | Full (linked device) |
| iMessage | macOS only (BlueBubbles bridge) |
| Microsoft Teams | Full (webhook + bot) |
| Google Chat | Full (app webhook) |
| Matrix | Full (homeserver bridge) |
| IRC | Full |
| Mattermost | Full |
| Nextcloud Talk | Full |
| Nostr | Full |
| Twitch | Full (chat bot) |
| Feishu / Lark | Full |
| LINE | Full |
| Zalo | Full |
| Tlon (Urbit) | Full |
| Synology Chat | Full |
| WebChat | Browser-based fallback |
You can run multiple platforms simultaneously. Each connection is configured in ~/.openclaw/config.yaml.
Inference Backends
OpenClaw is inference-agnostic. You set LLM_PROVIDER in your environment and switch without restarting.
Local GPU via Ollama
The most popular self-hosted option. Pull a model with ollama pull and point OpenClaw at the local endpoint.
# ~/.openclaw/config.yaml
llm:
provider: ollama
base_url: http://localhost:11434
model: llama3.3:70b
Model selection by VRAM:
| VRAM | Recommended Model | Quality |
|---|---|---|
| 6 GB | llama3.2:3b | Basic |
| 12 GB | llama3.2:latest (8B) | Good |
| 16 GB | llama3.1:8b-q8 | Better |
| 24 GB | mistral-nemo:12b | Strong |
| 48 GB | llama3.3:70b-q4 | Excellent |
| 94+ GB | llama3.1:405b-q2 | Near-frontier |
Cloud API Backends
For teams that do not need on-prem inference, OpenClaw works with any OpenAI-compatible API:
llm:
provider: anthropic
api_key: "${ANTHROPIC_API_KEY}"
model: claude-sonnet-4-5
Supported providers out of the box: Anthropic, OpenAI, Mistral, Groq, Together AI, Fireworks, and any OpenAI-compatible endpoint.
Local vLLM
For production-grade self-hosted inference at scale:
llm:
provider: vllm
base_url: http://localhost:8000/v1
model: meta-llama/Llama-3.3-70B-Instruct
vLLM supports tensor parallelism across multiple GPUs and achieves higher throughput than Ollama for concurrent sessions. That matters when multiple team members use the same OpenClaw instance.
Voice Support
OpenClaw supports voice input and output on macOS, iOS, and Android. The voice pipeline uses Whisper for transcription and a TTS engine for responses. Both can run locally (no API key required) or delegate to cloud APIs for higher quality.
- macOS: System tray app with push-to-talk
- iOS / Android: Official mobile app (OpenClaw app on App Store and Play Store)
Browser Control and System Automation
The agent runtime can control a real browser using Playwright. This enables tasks like:
- Filling forms and submitting data
- Reading and summarizing web pages
- Monitoring pages for changes
- Extracting structured data from sites without APIs
System automation capabilities include shell command execution, file operations, and HTTP requests. By default, these run inside a per-session Docker container. For NemoClaw deployments, the sandbox is hardened further with Landlock and seccomp.
ClawHub Skills
ClawHub is the official skill registry at clawhub.dev. Skills are npm packages that add domain-specific tools to the agent runtime.
Popular skill categories:
- Productivity: Google Calendar, Notion, Linear, Jira integration
- Data: Database queries, spreadsheet read/write, CSV parsing
- Communication: Email drafting, meeting summaries
- Development: GitHub PR summaries, CI status, code review
- Research: Web search aggregation, academic paper lookup
Install a skill from the dashboard or CLI:
openclaw skills install @clawhub/linear
ClawHub Supply Chain Risk
ClawHub's open contribution model is also its biggest security liability. Security researchers have identified 341 malicious skills in the registry: packages that pass as productivity tools but silently exfiltrate files or make outbound calls to attacker-controlled servers. Over 135,000 OpenClaw instances have been found reachable from the internet without authentication.
For teams handling sensitive data, this matters: any installed skill runs with the same process privileges as the OpenClaw agent itself. Base OpenClaw has no sandboxing layer to limit what a skill can access.
The mitigation is NemoClaw, NVIDIA's process-level sandbox that restricts filesystem access, blocks unauthorized network calls, and logs every action the agent takes. For healthcare and enterprise workloads, running NemoClaw alongside OpenClaw is the recommended baseline.
OpenClaw in 2026: Ownership and Independence
In February 2026, OpenClaw's creator Peter Steinberger joined OpenAI in a talent acquisition deal. OpenAI committed to keeping OpenClaw open-source under a foundation structure, with OpenAI as a sponsor rather than the owner. The project remains MIT licensed and community-maintained.
What this means practically:
- The open-source project continues. The codebase, ClawHub, and community contributions are unaffected. OpenAI's stated position is to accelerate the project, not absorb it.
- Cloud inference is not mandated. Nothing in the current roadmap forces users toward OpenAI-hosted inference. Self-hosted deployments using Ollama, vLLM, or NIM are fully supported and will remain so.
- Enterprise teams should plan for governance uncertainty. Any change in foundation structure or licensing would require migration. Teams with strict data sovereignty requirements should pin versions, maintain their own forks, and not depend on ClawHub for production-critical skills.
For context on why this acquisition accelerated NemoClaw's development, see What Is NemoClaw.
MCP (Model Context Protocol) Support
OpenClaw supports MCP servers as tool providers. This means any MCP-compatible server, including Apify MCP actors, can be connected as a tool source for the agent.
mcp:
servers:
- name: apify
url: https://mcp.apify.com
token: "${APIFY_TOKEN}"
This makes OpenClaw a useful orchestration layer: the assistant receives a task via Telegram, executes it using an Apify actor via MCP, and returns the result in the same conversation thread.
System Requirements
Minimum (cloud API backend, no local GPU)
| Resource | Requirement |
|---|---|
| CPU | 2 vCPU |
| RAM | 2 GB (4 GB recommended for builds) |
| Disk | 10 GB |
| OS | Ubuntu 22.04+ (macOS and Windows/WSL2 also supported) |
| Runtime | Node.js 22.16+ (Node 24 recommended), Docker Engine v24+ |
A 4 GB Liquid Web VPS ($8.50/mo on the current 50%-off-for-2-months promo, $17/mo list) comfortably runs the gateway and agent runtime with a cloud API backend.
Recommended (local Ollama, small model)
| Resource | Requirement |
|---|---|
| CPU | 4+ vCPU |
| RAM | 16 GB |
| GPU | 12+ GB VRAM (e.g., L4 Ada 24 GB on Liquid Web) |
| Disk | 40 GB NVMe |
Production (vLLM, large model, multi-user)
| Resource | Requirement |
|---|---|
| CPU | 16+ cores |
| RAM | 64–256 GB |
| GPU | 48–94 GB VRAM (L40S or H100) |
| Disk | 200 GB+ NVMe |
| OS | Ubuntu 22.04+ with NVIDIA drivers |
See OpenClaw GPU inference on LiquidWeb for a full hardware selection guide.
Versioning and Stability
OpenClaw uses calendar versioning (YYYY.M.patch). Releases ship multiple times per month with active community contribution (360+ contributors). There is no LTS branch, so always pin to a specific tag in production rather than latest. Check the releases page for the current stable tag before deploying.
# Replace the tag with the current release from GitHub
docker pull ghcr.io/openclaw/openclaw:2026.4.x
Alternatively, install via npm and run the onboarding wizard on the host:
npm install -g openclaw@latest
openclaw onboard --install-daemon
Yes. OpenClaw is MIT licensed and free to use, modify, and self-host. You pay for your own infrastructure (VPS, GPU server) and any cloud API keys you choose to use. There is no paid tier of OpenClaw itself.
Open WebUI is a chat interface for local models. OpenClaw is an autonomous agent: it can take actions (browser control, API calls, shell commands), connect to messaging platforms you already use (WhatsApp, Telegram, Slack), and execute multi-step tasks without you staying in a chat window.
Yes. Use any cloud API backend (Anthropic, OpenAI, Groq). A basic VPS with 2–4 GB RAM is sufficient. GPU is only needed for fully local inference.
ClawHub is the official skill registry for OpenClaw with 5,700+ installable skills. Skills are npm packages that add new tools to the agent, from Google Calendar to GitHub to SQL database queries.
Yes. OpenClaw can connect to any MCP server as a tool source. This includes Apify actors exposed via MCP, giving the agent access to web scraping and data extraction capabilities.
The project remains open-source under MIT license. OpenAI committed to a foundation structure where the project is community-maintained and OpenAI acts as a sponsor. In practice, the codebase, ClawHub, and self-hosting options are unchanged. Enterprise teams with strict data sovereignty requirements should pin specific versions and avoid depending on ClawHub for production-critical skills, as governance could evolve over time.
ChatGPT is a cloud service: your data is processed on OpenAI's infrastructure, and you use it through OpenAI's interface. OpenClaw is an autonomous agent you self-host. Your data stays on your hardware, you choose the inference backend (including local models with no cloud dependency), and the agent connects to messaging platforms you already use rather than requiring a new chat window.
OpenClaw is gateway-first: it routes messages to an LLM and executes skills in per-session containers. Hermes Agent is agent-first: it has persistent cross-session memory and a learning loop that builds reusable skill documents from experience. For teams that run the same workflows repeatedly, Hermes improves measurably over time. OpenClaw wins on messaging integration (10+ platforms), model flexibility, and lower resource requirements.
Related articles
- Docker Compose production stack: full Compose file with Redis, Caddy, and backups
- Skill development guide: build custom ClawHub skills
- Monitoring & observability: Prometheus metrics and alert runbook
- vs Alternatives: OpenClaw vs NemoClaw vs Hermes Agent vs Open WebUI
Common mistakes and fixes
OpenClaw dashboard is unreachable after install.
Confirm the container is running with 'docker ps'. The dashboard listens on port 18789 by default, so check that your firewall or security group allows it, or use an SSH tunnel: ssh -L 18789:localhost:18789 user@your-server.
Messaging platform integration fails to authenticate.
Each platform requires its own bot token or credentials. Re-read the platform-specific setup in the OpenClaw docs at docs.openclaw.dev. WhatsApp requires a QR scan on first link; tokens expire if the session is inactive.
