Lesson 01 · OpenClaw, Safely Free ~9 min Updated September 11, 2026

OpenClaw explained: a personal AI agent that lives on your computer.

250,000+ stars on GitHub, a lobster for a mascot, and a habit of showing up in your group chats. OpenClaw turned "AI assistant" from an app you open into an agent that messages you. Here's what it actually is, in plain English — including the parts the hype skips. Updated for OpenClaw 2.0, released August 30, 2026.

Version update · August 30, 2026

OpenClaw 2.0 changed how you start — and who can join in.

On August 30, 2026 the project shipped OpenClaw 2.0 (version 2026.8.1) — by its maintainers’ own account the largest update in OpenClaw’s history: 933 contributors, 569 of them first-timers, and over 16,000 pull requests, roughly half of every pull request ever merged into the project. The release touches installation, messaging, memory, skills, models, automations, the browser and native apps, plugins and security. It also broke a habit: OpenClaw had shipped 106 releases in 230 days, then went nearly seven weeks without one while the team reworked the foundation.

A follow-up, 2026.8.2, landed on September 1, 2026. The user-visible pieces: the Home agent can now sit in a right or bottom dock beside whatever you are working on (Cmd/Ctrl+Shift+H), a desktop companion arrived for x86-64 Linux (.deb or AppImage), background sessions start from New Session without leaving the page, and upgrades got safer — the updater now keeps newer config, refuses to declare a half-finished migration successful, and can recover a stopped Gateway after a failed update. Point releases like this are the normal cadence returning; 2.0 was the exception.

Two days later, 2026.9.1 (September 3, 2026) kept that cadence. The pieces a non-developer will notice: diagrams render in chat — when the agent answers with a Mermaid flowchart or org chart, the Control UI and the macOS, iOS and Android apps now draw it instead of showing code; a fresh install (npx openclaw@latest) can spot an existing Claude Code or Codex login and get you from install to first chat in one prompt, with the full wizard demoted to "Custom setup"; the Android app caught up with the web UI; and openclaw update now rolls itself back when its post-update health check fails, keeping your config and secrets intact. Every one of those is aimed at the same complaint this explainer has always carried: the setup and upkeep cost is the real price of OpenClaw, and it is coming down release by release.

The cadence held. 2026.9.2 (September 5, 2026) added GPT-6 Astra as a selectable model — openai/gpt-6-astra with an OpenAI API key or an eligible ChatGPT/Codex sign-in, where subscription access follows OpenAI’s own rollout (see who actually gets Astra) — and let most agent, model, tool and channel settings apply without restarting the Gateway. 2026.9.3 (September 8, 2026) brought a searchable meeting library (saved notes, full transcripts, Markdown or JSONL downloads from the Control UI), provider-account management inside Models settings, and a share link that publishes a revocable, read-only view of a conversation — including its future messages — to anyone holding the URL, so treat it like a public document, not a DM. The change most likely to bite: 2026.9.3 requires Node 24.16.0 or newer on 24.x, or Node 26.1.0 or newer (26 recommended); Node 22, Node 25 and older 24.x/26.x builds are no longer supported, and the project says to upgrade Node before OpenClaw to avoid SQLite text truncation.

2026.9.4 (September 11, 2026) is mostly upkeep, and the upkeep is the point: openclaw update can now roll back to the previous package and configuration when a compatible update fails its checks (database migrations still need the pre-update backup the lock-it-down lesson insists on); bundled and ClawHub plugins are discovered, installed and permissioned from a single Plugins workspace in the Control UI instead of three places; the agent can ask you multiple-choice questions in the terminal; and OpenAI’s new GPT Image 2.5 “Flare” and “Sunburst” variants are selectable for image work without changing your default model. A new OPENCLAW_CONFIG_READONLY=1 switch stops OpenClaw rewriting a config file you manage yourself.

Three changes matter if you are deciding whether to run one:

  • Setup got much shorter. A first-time install now starts from what is already on your computer — an existing ChatGPT or Claude subscription, API keys you already hold, models installed locally — instead of asking you to supply everything up front. Most configuration moved out of initial setup, so you reach a first conversation quickly and finish the rest by talking to the agent.
  • The browser app is now the front door. The web interface was rebuilt as a first-class experience rather than an afterthought, and it opens straight into a conversation with your agent. This is where most people will now meet OpenClaw.
  • Shared cloud sessions make it multiplayer. You can bring another person into a live session, or hand work over entirely, without the agent losing the context it already had. That is genuinely new: before 2.0, involving someone else meant starting over.
What this does not change

Easier setup is not the same as safer setup. The security trade below is exactly as it was, and shared sessions add a person who can now reach whatever your agent can reach — so the isolation choices matter more, not less. Work through the safe-setup checklist before you connect anything real, and see the update briefing for how the release cadence has been running.

The short version

ChatGPT is a place you visit. OpenClaw is a colleague who texts.

OpenClaw is a free, open-source program that runs on your own computer — not in a company's cloud. It wires an AI model you choose (Claude, GPT, others) into your real life: it reads and sends messages on WhatsApp, Telegram, Discord, Signal and a dozen more, browses the web, manages files, runs scheduled tasks, and acts on your behalf around the clock.

☁️ A cloud assistant

You open an app and start a conversation. It waits for you, in someone else's data center.

🦞 OpenClaw

Lives on your machine, reachable from the chat apps you already use, working on your schedule whether you're at a keyboard or not.

The trade in one line

OpenClaw gives you more power and more privacy than any cloud assistant — in exchange for making you the IT department. That trade is the whole story, and it's why the very next lesson is entirely about safety.

Do it · tell the pieces apart

Three names, three very different things.

People blur these together and get burned. Match each description to its name.

Name the piece0 / 3
Where it came from

Why the name keeps changing.

Late 2025
ClawdBot. Developer Peter Steinberger wires Claude into his Mac, talking to him over WhatsApp. It goes viral.
The rename
Moltbot. The community leans into the lobster joke — lobsters molt.
Jan 2026
OpenClaw. The name settles for good.
Feb 2026
Steinberger joins OpenAI. The project doesn't die — it's community-run, still shipping releases weekly.
Apr 2026
250,000+ GitHub stars — one of the fastest-growing open-source projects ever.
What it can actually do

It acts, it doesn't just answer.

💬
Meets you in your chat apps
WhatsApp, Telegram, Slack, Discord, Signal, iMessage, Teams — no new app, just another contact.
⚙️
Acts, not just answers
Browses the web, reads/writes files, runs shell commands, schedules jobs, responds to webhooks.
🧩
Learns via skills
Add-on instruction packs from ClawHub. Powerful and uncurated — a flea market, not an app store.
🧠
Brings your own brain
OpenClaw is the body; you plug in the model (Claude, GPT…) via your own API key. Swap anytime.
What it costs

Free software, real running costs.

🔑Model usage — it burns API tokens every time it thinks. You control it with model choice + limits.$ few → $20–100+/mo
🖥️An always-on machine — old laptop, Mac mini, Raspberry Pi, or a small cloud server.$0 → ~$10/mo
⏱️Your time — 15–30 min to set up, but ownership is ongoing; updates ship weekly.the honest line
Do it · is it for you?

Power and privacy, or none of the upkeep?

Mark each as you or not-you. The read updates as you go.

🦞 Fit check
Mark the statements above to see your read.
Hard stops

Two situations are flat "not yet," whatever your score: you planned to run it on your main work laptop next to client files and saved passwords (the single most common mistake), or your employer restricts unapproved software — several major companies have banned OpenClaw on work devices outright.

The part the hype skips

"Open source" does not mean "safe."

It means the code is public — including for attackers, who in early 2026 found tens of thousands of badly-configured OpenClaw instances exposed on the open internet. The difference between a private assistant and a public liability is entirely in how you set it up. That's not a footnote to this track; it is this track.

Where this is going

Microsoft's new always-on agent for Microsoft 365 — Scout, announced June 2026 — is built on the OpenClaw framework, wrapped in enterprise governance. The open-source lobster you can run tonight is the same foundation Microsoft chose for its enterprise autopilot. Learning OpenClaw is an early look at the architecture your workplace tools are adopting.

Before you install anything

Read the next lesson first. It's free, it's honest, and it covers what the hype videos don't: the exposed-instance scans, the malicious-skill problem, and a clear framework for whether you should run OpenClaw at all. Ten minutes there saves real grief later.

What you can do now

  • Explain what OpenClaw is: an open-source AI agent on your own machine, reachable via your chat apps
  • Tell the pieces apart: OpenClaw (the agent), ClawHub (the skills marketplace), Moltbook (an unaffiliated agent social network)
  • Name what it does: messaging, browsing, files, shell, schedules, webhooks, and add-on skills
  • Estimate the real costs: API usage, an always-on machine, and your maintenance time
  • Make the fit call: power + privacy + responsibility, or a cloud assistant with none of the upkeep
Free
Up next in OpenClaw, Safely

Lesson 02 · Is OpenClaw safe? The honest risk briefing

Tens of thousands of exposed instances. A poisoned skills marketplace. Corporate bans. What the security research actually found, and a clear framework for deciding whether to run it. Read it free →

🎓
AI Coach
Ask anything about this lesson
Hey! I'm your AI Coach for this lesson. Ask me anything about what you just read — what OpenClaw is, what it costs, whether it's a fit for you. What's on your mind?
Free lesson coaching is limited to 3 questions. Upgrade to Pro for unlimited coaching on every lesson.