Open source Β· Runs on your machine

The AI that really does things.

Organizes your inbox, sends emails, manages your calendar, checks you in for flights - all from WhatsApp, Telegram, or any chat app you already use. Your agent, your machine, your rules.

346k+GitHub stars
29chat channels
532contributors on v2026.7.1
MITlicensed, forever
β–Ά OpenClaw - quick demo & overview
🦞 2-minute demo - see the lobster do real work
~/quickstart - works on macOS, Linux, and Windows
What is OpenClaw?

Your own AI assistant. On your machine. In your chat apps.

OpenClaw is a free, open-source personal AI assistant that runs on your computer or server and actually does things - reads your email, manages your calendar, browses the web, writes code, runs commands - all from the messaging apps you already use. It started as a weekend hack in Austria and became the fastest-growing repository in GitHub history, now stewarded by the non-profit OpenClaw Foundation.

How does OpenClaw work?

One Gateway process runs on your machine and bridges everything: your chat apps on one side, an AI agent with tools, memory, and skills on the other. Message it from anywhere; it thinks with the model you chose and acts with the permissions you granted.

architecture - one gateway, everything connectedflow
you, from anywhere β”‚ β–Ό WhatsApp Β· Telegram Β· Discord Β· Slack Β· iMessage … 29 channels β”‚ β–Ό β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β” β”‚ GATEWAY - runs on your machine or server β”‚ β”‚ routing Β· sessions Β· permissions Β· memory β”‚ β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜ β”‚ β–Ό OpenClaw agent β”œβ”€ model Claude Β· GPT Β· Gemini Β· local LLM β”œβ”€ tools browser Β· shell Β· files Β· email Β· cal └─ skills ClawHub + the ones it writes itself

How to use it - just talk.

No dashboards to learn. After onboarding, you message your agent like a person. Real use cases from the community:

  • "Give me my morning brief - calendar, inbox, and anything urgent."
  • 3 meetings today. I moved the 2pm conflict, drafted a reply to the invoice email, and flagged one contract for your review.
  • "Unsubscribe me from all this newsletter spam."
  • "The tests are failing on main - run the fix loop and open a PR."
  • "Remind me to leave for pickleball based on traffic."
  • Done - I'll ping you ~20 minutes before, adjusted live.
#1fastest-growing repo in GitHub history
500k+systems running the lobster
6 mofrom weekend hack to Foundation
β˜…β˜…β˜…β˜…β˜…"iPhone moment," say early reviews
Latest - July 2026

Fresh from the reef.

Product notes with docs-level depth. The lobster ships fast.

Jul 13, 2026 Β· Release

v2026.7.1 is out

A revamped Control UI, conversational onboarding, GPT-5.6 / Claude Sonnet 5 / Tencent Hy3 / Muse Spark 1.1 support, ClawRouter, and stronger Codex workflows.

release
Jun 30, 2026 Β· Mobile

Now on iOS & Android

Native apps with chat, Talk mode, and remote action approvals - paired to your own Gateway, so your data stays yours.

apps
Jun 1, 2026 Β· Security

Skill security, with NVIDIA

Every ClawHub skill now ships with a Skill Card and is scanned by SkillSpector for hidden instructions and agentic risks.

security
Download & install

Get OpenClaw on any platform.

Windows, macOS, Linux, Docker, or a VPS - every OpenClaw installation path, from a one-line install to a fully self-hosted server setup.

macOS macOS 15+ Β· Universal binary (Apple Silicon & Intel)

Native app with controls for the gateway, chat, setup, and node features. First run may prompt for an administrator password for Homebrew.

⊞Windows Windows 10 20H2+ or Windows 11 · x64 & ARM64

The Windows Hub runs OpenClaw with tray, setup, chat, and node mode - natively contained via Microsoft Execution Containers, straight from the Build 2026 keynote.

>_Linux Any modern distro Β· Node.js 22+ (auto-installed)

Install OpenClaw on Linux with the one-liner below - it handles Node.js and every dependency for you. Ideal for servers, homelabs, and a Raspberry Pi 5.

One-liner

curl -fsSL https://openclaw.ai/install.sh | bash

The fastest way to install OpenClaw on macOS and Linux. Installs Node.js and everything else, then run openclaw onboard.

npm install

npm i -g openclaw

Already have Node.js 22+? Install OpenClaw globally via npm and onboard from any terminal on any OS.

Docker

docker Β· container setup

Run the Gateway in a container for clean isolation on servers, NAS boxes (TrueNAS, Portainer), and homelabs. Follow the Docker guide in the docs.

From source

git clone github.com/openclaw/openclaw

The hackable install: clone, pnpm install, and run from source. Best for contributors and plugin developers.

Self-hosted setup: host OpenClaw on a server

Want a 24/7 assistant that answers even when your laptop sleeps? Run OpenClaw on a small VPS or any always-on machine you own. It's the same install - just on a server.

min spec: 2 vCPU Β· 4 GB RAM Β· Node.js 22+ Β· ~$4–10/mo
  1. Provision a VPS - any small Linux server works (Hetzner, DigitalOcean, Oracle's free ARM tier, or your own hardware).
  2. Run the installer - curl -fsSL https://openclaw.ai/install.sh | bash sets up Node.js and OpenClaw in one go.
  3. Onboard your lobster - openclaw onboard walks you through model providers and credentials conversationally.
  4. Connect a channel - link Telegram, WhatsApp, Discord, or any of the 29 channels, and confirm the Gateway isn't publicly exposed.
  5. Lock it down & go - set spend caps, keep API keys tidy, and update anytime with openclaw update. Pair the iOS/Android app for approvals on the move.
What it does

A real assistant, not a chat window.

[~/local]

Runs on your machine

Mac, Windows, or Linux. Bring hosted, subscription-backed, gateway, or fully local models. State lives with you, not a vendor cloud.

[29 ch]

Chats where you are

WhatsApp, Telegram, Discord, Slack, Signal, iMessage, and more - in DMs and group chats alike.

[mem++]

Persistent memory

Remembers your preferences, projects, and people from one conversation to the next. Context that persists 24/7.

[www→]

Browser control

Browses the web, fills forms, and extracts data from any site - checked in for your flight before you land at the gate.

[$ sh]

Full system access

Read and write files, run shell commands, execute scripts. Full access or sandboxed - always your choice.

[+skill]

Skills & plugins

Extend with community skills from ClawHub or build your own. It can even write its own - and hot-reload them.

OpenClaw skills & plugins

Teach your lobster new tricks.

Tools are callable actions, skills teach repeatable workflows, and plugins add runtime capabilities like providers and channels. Together they turn a chat agent into a workflow automation engine - and OpenClaw can even write its own.

ClawHub - the skill library

The public registry for reusable skills and installable plugins and extensions, built by the community. Install anything in one command or straight from a chat message.

πŸ›‘ Every ClawHub skill ships with a Skill Card and is scanned by SkillSpector for hidden instructions - the skill-security pipeline built with NVIDIA.
Browse ClawHub

Email skills

inbox

Inbox triage, smart replies, bulk unsubscribing, and a morning brief of what actually matters. Users report clearing 10,000-email backlogs on autopilot.

triage Β· draft Β· unsubscribe Β· brief

Calendar skills

schedule

Timeblock tasks, resolve conflicts, get briefed before meetings, and get pinged when it's time to leave based on live traffic.

timeblock Β· conflicts Β· pre-meeting briefs

Coding agent

code

Delegate to Claude Code or Codex from any chat: run autonomous fix-the-tests loops, steer sessions from your phone, and attach to running work with openclaw attach.

Claude Code Β· Codex Β· /autoreview Β· /handoff

GitHub issues & PRs

devops

Create and triage issues from chat, watch webhooks (a Sentry error can become a fix and an open PR), and preview issue status inline in the Control UI.

issues Β· PRs Β· webhook-triggered fixes

CRM & lead generation

business

Connect CRMs like HubSpot via plugins or MCP: look up contacts, track deals, log activity, research prospects, and draft outreach - with you approving what goes out.

contacts Β· deals Β· research Β· outreach drafts

Business & task automation

ops

Cron jobs, heartbeats, reminders, and background tasks: expense reports, invoices, meeting-notes reviews, ad optimization, content pipelines - one user says "it's running my company."

cron Β· webhooks Β· background tasks
01 Β· Install

Grab a skill from ClawHub

Thousands of community skills and plugins, one command away - or ask your agent to install one mid-conversation.

"install the todoist skill" β†’ done
02 Β· Ask

Let the agent build its own

Describe the workflow in chat and OpenClaw writes the skill itself, hot-reloads it, and starts using it - no editor required.

"build a skill for my uni assignments" β†’ built
03 Β· Refine

Promote work with Skill Workshop

The Skill Workshop turns completed agent work into reusable skills - you review, revise, and approve proposals before they change how your agents operate.

review β†’ revise β†’ apply

The result: a productivity assistant that compounds - every workflow you teach it once, it runs forever. Skills docs β†—

Pricing

How much does OpenClaw cost?

OpenClaw is free, open-source software - no license fee, no seats, no subscription. What you pay for is where it runs and which AI model powers it. Here's the honest breakdown.

$0

The software is free. Forever.

OpenClaw is MIT-licensed with no paid tiers, no OpenClaw cloud subscription, and no usage caps from the project itself. Your real monthly cost = hosting + model/API usage, and both start at zero.

Fully local $0/mo old laptop Β· Mac mini Β· Pi 5
  • Run on hardware you already own
  • Local models (Ollama etc.) - no API bill
  • Free-tier cloud (e.g. Oracle ARM) also works
  • Best for tinkerers & privacy-first setups
Typical personal $10–40/mo most users land here
  • VPS: ~$4–10/mo (2 vCPU / 4 GB is plenty)
  • API usage: ~$5–30/mo at moderate load
  • Or reuse a Claude / ChatGPT subscription
  • 24/7 assistant, always reachable
Power user / team $50–150/mo heavy automation
  • Thousands of multi-step workflows
  • Browser sessions & complex reasoning
  • Multi-agent fleets across machines
  • Smart model routing keeps this in check
Managed hosting $29–55/mo third-party services
  • Community-run hosts handle setup & uptime
  • Some bundle tokens, some are BYO-key
  • Not run by the OpenClaw Foundation
  • Compare what's actually included
API & token costs - what actually drives the bill
ModelGood forper 1M tokens (in / out)
Claude Haiku Β· Anthropicheartbeats, routing, quick tasks$1 / $5
Claude Sonnet Β· Anthropicthe best default for most work$3 / $15
Claude Opus Β· Anthropicgenuinely hard reasoning only$5 / $25
GPT-5 Nano / Mini Β· OpenAIcheap high-volume chatfractions of a cent per msg
GPT-OSS-120B Β· open-weightbest value tool-calling$0.04 / $0.10
Local models Β· Ollama etc.zero-API setups$0

# A typical OpenClaw task uses 2,000–8,000 tokens across 3–8 model calls. Rates shift - check your provider's current pricing. You can also sign in with a Claude or ChatGPT subscription instead of paying per token.

Is there an OpenClaw subscription?

No. OpenClaw has no paid plans and no official cloud - it's a free, MIT-licensed OpenClaw Foundation project. Any "OpenClaw cloud pricing" you see comes from independent third-party hosts, not the project itself.

What does self-hosting cost?

$0 on hardware you already own, or roughly $4–10/month for a small VPS (2 vCPU / 4 GB RAM is enough). Add your model spend on top, plus a couple of hours of setup and occasional maintenance.

What will I spend on Claude or OpenAI APIs?

Light use runs under $1–5/month; moderate personal use around $10–30; heavy automation $50–150+. Model choice matters more than server size - route routine tasks to cheap models and reserve premium ones for hard reasoning. Alternatively, connect an existing Claude or ChatGPT subscription.

The #1 cost mistake to avoid?

Leaving the heartbeat (the periodic "anything to do?" wake-up) pinned to a premium model. Point it at a cheap model or lower its cadence, and an idle agent costs cents instead of dollars per day.

OpenClaw API & models

Bring any brain. Claude, GPT, Gemini, or your own.

OpenClaw separates model choice, authentication, and execution runtime. Connect with an API key, sign in with a subscription you already pay for, route through a gateway, or run models fully local - and switch anytime with one config change.

~/.openclaw/openclaw.json - model configurationjson
{
  "models": {
    // your main agent - pick the strongest model you can
    "default": "anthropic/claude-sonnet",
    // cheap model for titles, summaries & background chores
    "utilityModel": "google/gemini-flash",
    // don't let an idle agent bill premium tokens
    "heartbeat": "anthropic/claude-haiku"
  }
}

# Illustrative - see the model configuration docs for exact keys, aliases, and per-agent overrides. openclaw doctor preserves your model tuning across updates.

Best model for OpenClaw? Route, don't choose.

There's no single best model - the winning setup routes each job to the cheapest model that does it well. A sensible policy cuts the model bill 40–70%:

  • β†’Daily driver: Claude Sonnet or GPT-5.x - the best balance of capability and cost for most agent work.
  • β†’Hard reasoning only: Claude Opus or a frontier reasoning model, reserved for genuinely complex tasks.
  • β†’Background chores: Haiku or Gemini Flash for heartbeats, titles, classification, and short summaries.
  • β†’Privacy-critical: a local model via Ollama - slower, but nothing leaves your machine.

API keys stay on your machine in your own config - OpenClaw is the harness, not a middleman.

OpenClaw integrations

One agent. Your whole stack.

29 chat channels plus tools for email, calendar, code, and the open web. Every channel installs in one command or during onboarding - run several at once and OpenClaw routes each conversation automatically.

Chat channels - talk to your agent anywhere
Tools & automation - what your agent can do

+ Matrix, Mattermost, IRC, LINE, Feishu, QQ, Nextcloud Talk, SMS, Voice Call, Twitch, and more - browse the full directory β†—

β—†WhatsApp Β· channel β—†Telegram Β· channel β—†Discord Β· channel β—†Slack Β· channel β—†iMessage Β· channel β—†Signal Β· channel β—‡Claude Β· model β—‡GPT Β· model β—‹Browser Β· web β—‹Gmail Β· inbox β—‹GitHub Β· code β—‹Obsidian Β· notes β—‹Spotify Β· tool β—‹Philips Hue Β· home β—‹Twitter / X Β· web β—†WhatsApp Β· channel β—†Telegram Β· channel β—†Discord Β· channel β—†Slack Β· channel β—†iMessage Β· channel β—†Signal Β· channel β—‡Claude Β· model β—‡GPT Β· model β—‹Browser Β· web β—‹Gmail Β· inbox β—‹GitHub Β· code β—‹Obsidian Β· notes β—‹Spotify Β· tool β—‹Philips Hue Β· home β—‹Twitter / X Β· web
OpenClaw security & privacy

Is OpenClaw safe? Here's the honest answer.

OpenClaw is powerful software - it can read files, run commands, send messages, and act on schedules. That's the point, and it's also why security is a shared responsibility: OpenClaw ships safe defaults, and you decide what it can touch. Treat it like privileged software, not a harmless chatbot.

Local data & privacy

privacy

Your context, memory, skills, and state live on your machine - not a vendor cloud. There's no hosted service in the loop, no telemetry harvesting your conversations, nothing to leak from someone else's server.

self-hosted Β· your hardware Β· your data

Sandbox & isolation

sandbox

Run the agent with full access or sandboxed - your choice. On Windows, OpenClaw runs inside Microsoft Execution Containers; on servers, Docker gives clean isolation. Unsafe runtimes are stopped before they can touch state databases.

execution containers Β· docker Β· scoped runtimes

Permissions & pairing

access

Channels default to pairing and allowlists: unknown senders are ignored until you approve them, groups are blocked unless allowlisted, and sensitive node capabilities (like browser proxying) require admin-level approval.

allowlists Β· DM pairing Β· operator roles

Skill security

supply chain

Community skills are a supply chain, and OpenClaw treats them like one: every ClawHub skill ships a Skill Card documenting what it does and where it came from, and is scanned by SkillSpector for hidden instructions and agentic risks.

skill cards Β· SkillSpector Β· NVIDIA collab

Credential security

secrets

API keys and channel credentials live in your local config with masked prompts during onboarding - OpenClaw is the harness, not a middleman. Keep secrets out of version control and rotate anything you paste into a chat.

local config Β· masked prompts Β· your keys

Human-in-the-loop

approvals

Remote action approvals on the iOS and Android apps, explicit approval binding during setup, and spend caps mean the agent asks before consequential actions - you stay the operator.

action approvals Β· spend caps Β· audit your logs

⚠Know the real risks: prompt injection & over-permissioning

Like every agentic AI system, OpenClaw can be targeted by prompt injection - malicious instructions hidden in emails, web pages, or skills that try to steer the agent. No harness fully solves this today. The practical defense is least privilege: give the agent narrow jobs, separate accounts, limited permissions, and human review for anything touching money, reputation, clients, or confidential files. An open-source, self-hosted agent means you control the blast radius - but "local" is not the same as "safe." Boring, disciplined setups win.

Secure OpenClaw setup checklist

The hardening steps that matter most, whether you run it on a laptop or host it on a server. Ten minutes of discipline now saves a bad week later.

defaults are safe Β· these keep them that way
  1. Don't expose the Gateway publicly - bind it locally and reach it over Tailscale, a VPN, or SSH instead of an open port.
  2. Use explicit allowlists - list who may message the agent per channel; never ship allowFrom: ["*"] to production, and require mentions in groups.
  3. Sandbox and scope - run in a container where possible, give the agent its own accounts (not your primary logins), and grant the minimum scopes each integration needs.
  4. Guard credentials & spend - keep keys in local config only, set provider spend caps, and review the agent's memory and logs periodically.
  5. Vet skills & stay updated - install from ClawHub with Skill Cards, review skills before applying them, and keep OpenClaw current with openclaw update for security fixes.

Security posture evolves fast in agentic AI - read the current guidance in the security docs β†— and the security blog posts β†—.

OpenClaw alternatives & comparisons

How OpenClaw compares.

"Best alternative" depends on what you're actually replacing. Most tools people compare with OpenClaw solve a different problem - here's a fair map of the landscape, grouped by what each is for.

vs ChatGPT & Claude

hosted chat

ChatGPT and Claude are polished hosted assistants you visit. OpenClaw is a self-hosted agent that lives in your chat apps 24/7 with persistent memory and system access. Not really rivals: many users plug their ChatGPT or Claude subscription in as OpenClaw's brain.

different layer - OpenClaw is the harness, they're the models

vs Claude Cowork & Manus AI

agent apps

Claude Cowork (Anthropic's agentic desktop app) and Manus are managed agent products - smoother onboarding, vendor-run, vendor-priced. OpenClaw trades that polish for open source, full self-hosting, any model, and 29 chat channels.

managed convenience vs open control

vs AutoGPT, CrewAI & LangChain

frameworks

These are developer frameworks for building agents - libraries, not products. OpenClaw is a ready-to-run assistant out of the box. If you want to ship your own agent system, use a framework; if you want an assistant today, run OpenClaw.

build-your-own vs ready-to-run

vs n8n, Zapier & Make

workflow tools

Trigger-action platforms excel at deterministic pipelines: when X happens, do Y, every time. OpenClaw handles open-ended, conversational work that doesn't fit a flowchart. They pair well - plenty of setups have OpenClaw triggering n8n flows and vice versa.

deterministic pipelines vs open-ended agency

vs Lindy AI & Relevance AI

saas agents

Hosted AI-assistant platforms with per-seat or usage pricing, built for teams that want zero infrastructure. OpenClaw is free and self-hosted: you own the data and the bill, but also the setup and maintenance.

subscription SaaS vs free self-hosted

vs Hermes Agent & NanoClaw

closest peers

The nearest true alternatives: other personal agent harnesses. Hermes is an alternative stack (some users run both - their agents even chat), and NanoClaw is a security-focused fork-alike with stricter human-in-the-loop defaults.

same category - pick your philosophy
At a glance - what actually differs
OpenClawHosted assistants (ChatGPT, Claude, Manus, Lindy)Frameworks (AutoGPT, CrewAI, LangChain)Workflow tools (n8n, Zapier, Make)
Open sourceβœ“ MITmostly βœ—βœ“ mostlyvaries
Self-hosted / local dataβœ“ alwaysβœ— vendor cloudβœ“ you build itn8n βœ“ Β· others βœ—
Lives in your chat appsβœ“ 29 channelsown app / limitedif you build itβœ— triggers only
Model choiceβœ“ any, incl. localvendor's modelsβœ“ anylimited
Ready to use todayβœ“ productβœ“ productβœ— libraryβœ“ product
Software cost$0subscription$0 + dev timefree tier β†’ paid

# Simplified by necessity - each column groups distinct products. The honest takeaway: OpenClaw's edge is open, self-hosted, chat-native agency; managed products' edge is polish and zero maintenance. Choose for your constraints, not the hype.

Is OpenClaw worth it?

Worth it? Depends on who's asking.

The honest cost-benefit: the software is free, but you invest setup time, a small hosting bill, model costs, and the discipline to run privileged software safely. Here's who comes out ahead - and who doesn't.

βœ“Worth it if you…

  • Want an assistant that acts, not just chats - inbox triage, calendar juggling, code fixes, and web errands running 24/7, reachable from any chat app.
  • Already pay for Claude or ChatGPT - plugging an existing subscription in as the brain makes the marginal cost close to zero.
  • Care where your data lives - memory, context, and credentials stay on hardware you control, not a vendor cloud.
  • Are technical or curious enough to tinker - comfortable with a terminal, or willing to spend an evening learning. The payoff compounds: every skill you teach it runs forever.
  • Value open source - MIT-licensed, Foundation-stewarded, hackable down to the prompt, with a huge community shipping skills weekly.

βœ—Skip it (for now) if you…

  • Want zero setup and zero maintenance - a hosted assistant or managed agent app will serve you better than owning updates and config.
  • Only need deterministic automations - "when X, do Y" pipelines are cheaper and more reliable in Zapier, Make, or n8n.
  • Can't give it safe boundaries - if you'd hand it your primary accounts without allowlists, spend caps, or review, the risk outweighs the magic.
  • Need enterprise guarantees today - SLAs, compliance certifications, and vendor support are the domain of commercial platforms, not a young open-source project.
  • Expect polish over power - it improves at a wild pace, but you're an operator of fast-moving software, not a consumer of a finished product.
VERDICT 🦞

For the right person, emphatically yes. If you're technical, privacy-minded, or simply curious, OpenClaw delivers something no hosted product does: a personal agent you fully own, for roughly $0–40/month and an evening of setup. Half a million systems and the loudest word-of-mouth in AI - "iPhone moment," "it's running my company" - didn't happen by accident. Start small: one channel, one low-risk workflow, spend caps on. If the magic lands, expand. If it doesn't, you've lost an evening and a few dollars - the software itself never costs a cent.

From the community

What people are saying.

The fastest-growing project in GitHub history - a weekend hack that became the most-starred software repo in under five months, built by one creator, dozens of maintainers, and thousands of contributors.

- as told at the Y Combinator community celebration

Users describe daily briefings, inbox triage, calendar juggling, expense reports, deploy fixes over voice notes, and multi-agent fleets running across several machines - all steered from a phone chat.

- community showcase threads

β€œIt's running my company.”

- @therno, one of thousands of daily shoutouts
Get started

Meet your lobster in two minutes.

The one-liner installs Node.js and everything else for you. Switch channels later with a single command.

curl -fsSL https://openclaw.ai/install.sh | bash
View on GitHub Join the Discord Docs