How to Use OpenClaw: The Beginner's Guide (First Install, First Agent, First Message, and the Sandbox-First Rule)

CodexFuse Team · 2026-10-31T08:29:07.171Z · 14 min read min read

OpenClaw runs your own AI assistant on your machine. This beginner guide covers install, onboarding, gateway checks, model keys, Telegram, and the sandbox-first rule.

How to Use OpenClaw: The Beginner's Guide (First Install, First Agent, First Message, and the Sandbox-First Rule)

Learning how to use OpenClaw is the moment your AI assistant stops living inside a browser tab and starts living on your own machine. OpenClaw is a free, open-source AI assistant that runs on your computer or server and meets you in the chat apps you already use — Telegram, WhatsApp, Discord, Slack, and more than twenty others — so you message it like a person and it does real work: running tools, reading files, and taking actions on your behalf. This OpenClaw tutorial walks you through the complete beginner arc in one guide: installing it in one command, running the onboarding wizard, verifying the Gateway is healthy, connecting your first model, sending your first message from the dashboard, wiring up Telegram, and the sandbox-first safety rule that keeps an always-on agent from doing something you will regret. The cluster this post belongs to is the AI-agents cluster — its sibling lane is Aider beginner guide, the session-based pair-programming lane, and this page fills the cluster's missing persistent-agent lane: OpenClaw is the agent that stays running while Aider is the assistant you open for one coding session.

What is OpenClaw?

OpenClaw is an open-source personal AI assistant that runs on your own device instead of someone else's cloud: a local control layer called the Gateway connects your chat apps, a model, tools, memory, and scheduled jobs, so the assistant is reachable from your phone at any hour without a vendor's hosted service in the middle. It is stewarded by the OpenClaw Foundation, an independent nonprofit, it carries an MIT license, and the project explicitly has no paid tier, no hosted service, and no token — you bring your own model API key, and everything else is yours.

The project has had three names in less than a year, which is why beginner guides disagree with each other: Austrian developer Peter Steinberger published it in November 2025 as Clawdbot, it was renamed Moltbot, and it finally became OpenClaw. Anything in an old guide that says "clawdbot" or "moltbot" is the same software wearing an older name. The growth explains why it keeps appearing in your feeds: the repository crossed 134,000 stars in its first month, passed 335,000 by March 2026 to become the most-starred project on GitHub, and sat near 393,000 stars in late October 2026. The five-step arc from zero to a working agent is short enough to do in one sitting:

1. Install — one command installs the runtime and the openclaw CLI (about two minutes).
2. Onboard — the wizard verifies model access, creates your workspace, and configures the Gateway.
3. Verify — openclaw gateway status confirms the Gateway is healthy.
4. Message — openclaw dashboard opens the Control UI; send one test message there.
5. Connect a channel — link one messaging app (Telegram is the easiest) so the agent leaves the browser and lives on your phone.

What do you need before you start?

Keep the stack deliberately small for your first setup: one machine to run OpenClaw (macOS, Linux, Windows, or WSL2 — Windows users should run it under WSL2), one model provider with valid credentials, and one interface for the first test — the web dashboard the installer opens for you. A messaging account is optional at the start and recommended the moment the basic loop works.

The model is the one thing OpenClaw is not: OpenClaw adds the workspace, tools, memory, schedules, and communication channels, but the intelligence comes from a model you supply — Claude, OpenAI, or a local model through Ollama. Do not start with your main production API key. The writer's rule for this guide is simple: use a spending-capped key for onboarding, keep it out of shell history, and never paste it into a public chat. Fifteen to twenty minutes is enough if you do not skip the Node check below.

How do you install OpenClaw?

The official README offers three install paths that all end in the same place; pick exactly one. The one-line installer is the recommended path on macOS, Linux, and WSL2 because it detects your OS, provisions a supported Node.js runtime if you do not have one, and launches onboarding automatically on a fresh install. Before you run anything that pipes to bash, read the script URL once in your browser so you know what it does — a one-line install script is a program you are about to execute with your user privileges.

If you took the npm route — or if the installer did not open the wizard — start it yourself with openclaw onboard --install-daemon. The --install-daemon flag is what makes the Gateway a background service: it registers the Gateway to start on every boot, which is what turns OpenClaw from a program you launch into an assistant that is simply there. Verify the install with openclaw --version before continuing; if the command is not found, your global npm bin path is not on your PATH — fix that first instead of working around it.

Docker and Nix setups are documented in the official docs for containerized or reproducible installs — this guide uses the direct install path because the beginner arc is short, and the Docker beginner guide is there if the container branch is the one you want. If your Gateway will live on a machine that is always on, a server beats a laptop — the what is VPS hosting explainer covers why.

How does the onboard wizard work?

The onboarding wizard is where OpenClaw asks you the four questions that define your setup, and it is also where most beginners accidentally misconfigure security or wonder later why channels do not respond — so do not rush it. First it verifies access to a working model (you supply the provider credentials you prepared earlier), then it creates your workspace (the default folder where the assistant's memory, skills, and files live), then it configures the Gateway, and finally it offers the daemon install that keeps the Gateway available in the background.

Keep the first pass deliberately boring: one model, the default workspace, local access, and no optional skills until the basic loop works. Everything the wizard writes lands in a single configuration file, usually ~/.openclaw/openclaw.json, which you can inspect and edit afterward — nothing is hidden. Onboarding finishes by opening the Control UI in your browser, which is your next stop.

How do you check the gateway is running?

The Gateway is OpenClaw's always-on control layer: it is the process that connects the dashboard, the model, your messaging channels, the tools, sessions, and scheduled jobs, and every command you learn later assumes it is healthy. Check it with openclaw gateway status. If it reports healthy, open the dashboard with openclaw dashboard and send one short test message right there in the Control UI: ask it to reply with the current date and the model it is using.

That message is your first real conversation, and it is also your first diagnostic tool. If the agent answers with the date and names its model, your install, onboarding, model credentials, and Gateway are all working. If it does not answer, run openclaw doctor and check three things in order: model credentials, network access, and the Gateway logs. Do not add channels, skills, or scheduled jobs until this single message works — a broken basic loop cannot be fixed by adding more surface area to it.

Which model should you connect first?

Connect one model and only one for the first week. Cloud APIs (Claude, OpenAI) are the smoothest start because there is nothing to host, but they bill per token — set a spending cap in the provider's console before you onboard, and treat the key like a password. Most personal use costs only a few dollars a month, but an always-on agent with scheduled jobs can surprise you if you never set a cap.

The free alternative is running a model locally with Ollama: pull a model, point OpenClaw at it as the provider, and your assistant costs nothing per message and your conversations never leave your machine. The tradeoff is quality and resources — local models need a machine with enough memory, and smaller models follow long instructions less reliably than the cloud flagships. The Ollama beginner guide covers getting the local side running. Start with one lane (cloud or local), prove the loop works, and only then experiment with the other.

How do you talk to OpenClaw from Telegram?

Connecting one messaging channel is the step that turns OpenClaw from a dashboard demo into an assistant that lives in your pocket — and the beginner rule is one channel first, never three at once. Each new channel is another surface where misconfigured permissions can leak, and debugging is miserable when you cannot tell whether a message came from Telegram, Discord, or the dashboard. Telegram is the easiest first channel: create a bot through BotFather, take the token it gives you, and pair it in OpenClaw's channel setup so messages you send to the bot are routed through the Gateway to your agent and the replies stream back in-chat.

Once the pairing works, send the same test you sent in the dashboard — ask for the current date and the model name — and confirm the answer arrives on your phone. That is your second diagnostic: dashboard works but Telegram does not means a channel problem, not a model problem. Resist the urge to add WhatsApp, Discord, and Slack in the same evening. One channel, one week, one working habit — the community's most common beginner mistake is wiring five channels on day one and spending day two unable to tell which integration broke.

What is SOUL.md, and how do skills work?

OpenClaw agents are configured with a plain Markdown file called SOUL.md — it holds the agent's personality, role, instructions, and behavior rules in one human-readable document, instead of Python classes or YAML configs. If you can write a text document, you can give your agent a job: name it, describe what it should do, list what it must never do, and the Gateway enforces it. This is the single biggest difference from node-editor agent frameworks — your agent's brain is a file you can read and edit.

Skills are how the agent learns new abilities: reusable packages (written in JavaScript or Python, or plain Markdown for simple ones) that teach it how to do things like scrape a page, query an API, or run a browser flow. Community templates cover common roles — SEO analyst, content writer, DevOps monitor, and many more — so beginners can start from a template and read its SKILL.md to learn the format. Install new skills one at a time, verify each one with a test task before adding the next, and remember that a skill is code with permissions: only install skills from sources you trust, because a malicious skill runs inside your assistant with your tools.

What is the sandbox-first rule?

The sandbox-first rule is the safety habit that keeps an always-on agent from becoming a liability: OpenClaw can take real actions — run shell commands, modify files, call APIs, send messages — and a confident-sounding agent is not the same as a correct one. So your first jobs must be read-only: summaries, reports, drafts, and research that you review before anything is sent or changed. The three-part formula the community recommends for the first real job is trigger + task + boundary: "At 8:30 a.m., summarize my open tasks and suggest next actions, but do not modify anything." Only when the result is dependable do you schedule it to run automatically.

The two hard security rules are non-negotiable. First, the Control UI is an admin surface — keep the Gateway bound to your local machine by default and never expose the dashboard to the public internet without authentication in front of it, because an open Control UI is a remote-control panel for everything your agent can do. Second, keys are keys: the Gateway's tokens and your model API keys live in the configuration file, so that file gets strict permissions and never goes into a public repository. Self-hosting is the whole point of OpenClaw — and self-hosting means you are the one carrying the risk profile, not a vendor.

OpenClaw vs Aider: which should a beginner pick?

This is the comparison beginners actually need, because OpenClaw and Aider sit in the same AI-agents cluster but solve different problems — and the queue's difficulty evidence says the field has not settled this question for them. The answer is a lane decision, not a quality verdict:

They are not rivals: a beginner can reasonably run both — Aider for focused coding sessions, OpenClaw for everything the coding session cannot see. If the pair-programming lane is what you need today, the Aider beginner guide covers it.

Beginner mistakes to avoid

Every mistake below comes from the community's beginner guides and setup stories — they are the traps that turn a 15-minute install into a lost weekend:

Sources

• OpenClaw README (github.com/openclaw/openclaw) — install paths, quick start (onboard, gateway status, dashboard), MIT license, OpenClaw Foundation stewardship, via a 1-day-old crawl of the repository page.
• OpenClaw official docs (docs.openclaw.ai) — getting-started and configuration paths, referenced via an independent tutorial crawled 2026-10-31.
• ayautomate.com "OpenClaw GitHub Repo" (7 days ago) — 392,885 stars on October 24, 2026; release v2026.10.6 dated October 22, 2026.
• edrawmind.wondershare.com OpenClaw installation guide (15 hours ago) — Peter Steinberger, November 2025 Clawdbot launch; 335,000 stars by March 2026; Node 22 minimum; 15–20 minute setup.
• aiworkflows.tools "OpenClaw Complete Guide 2026" (updated September 8, 2026) — 134,000+ stars in the first month.
• clawdir.com "Getting Started with OpenClaw in 2026" (265 days ago) — five-step setup arc, Telegram channel flow.
• hackernoon.com "The Complete OpenClaw Setup Guide" (231 days ago) — two-track install-then-secure format, Gateway lockdown.
• myclaw.ai "OpenClaw Getting Started" (273 days ago) — one-model-one-interface start, trigger + task + boundary job formula.
• oneclaw.net "The OpenClaw README, Explained" (69 days ago) — README install-path decode, Gateway/Channels/Tools/Model-providers vocabulary.
• crewclaw.com "How to Use OpenClaw: Complete Beginner Guide (2026)" — SOUL.md concept, community skill templates.

Frequently asked questions

Is OpenClaw free to use?

Yes — OpenClaw itself is free and MIT-licensed, with no paid tier, hosted service, or token on the project side. The one cost is the model: cloud API usage (Claude, OpenAI) bills per token, typically a few dollars a month for personal use, and local models through Ollama are free after the hardware. Set a spending cap before you onboard.

What happened to Clawdbot and Moltbot?

They are the same project under earlier names. Peter Steinberger published it as Clawdbot in November 2025, it was renamed Moltbot, and it is now OpenClaw. If an old guide's commands reference "clawdbot" or "moltbot," the concepts transfer but the exact commands may not — verify against the current README.

Can I use OpenClaw without an API key?

Yes, with a local model through Ollama — pull a model, set Ollama as the provider, and your assistant runs with zero per-message cost and nothing leaving your machine. The tradeoff is resources: your machine needs enough memory, and smaller local models are less reliable at long instructions than the cloud flagships.

Is it safe to expose the OpenClaw dashboard on the internet?

Not by itself. The Control UI is an admin surface that can drive everything your agent can do, so keep the Gateway bound to your local machine and never put the dashboard on the public internet without authentication in front of it. The safest beginner posture is simple: local Gateway, token auth on the config file, and read-only jobs until you trust the loop.

Does OpenClaw work on Windows?

Yes — either through the PowerShell installer for native Windows, or by running it under WSL2 with the Linux installer. The community recommends WSL2 for the most-tested path, and there are native apps for Windows, macOS, iOS, Android, and Linux.

OpenClaw's pitch is simple: your assistant should live where you live, not in somebody else's browser tab. Install it once, prove the loop with one dashboard message, go read-only until you trust it — and then it is the assistant that is still awake when you wake up.