OpenClaw: Our Comprehensive Guide for Beginners

"I have my Claw. Now what?"

This is the question everyone asks as soon as they get their Claw. You now have an infinitely patient, extensible assistant that can talk, use tools, and run recurring jobs. But having infinite options can be as intimidating as having none.

This guide exists to help you understand what a Claw can do for you, and get specific ideas to integrate into your work and life so that it feels like magic without losing sight of the setup, permissions, and maintenance involved.

What's a Claw?

Your claw is a personal AI assistant that lives in whatever messaging app you choose—WhatsApp, Telegram, Discord, or SMS. You talk to it the same way you'd text a friend, once you've installed OpenClaw and connected a channel. Under the hood, it's powered by the same kind of AI behind tools like Claude and ChatGPT, running through OpenClaw—an open-source framework for personal AI assistants created by Peter Steinberger.

Here's what's most important: It has a personality, it can operate 24/7 as long as its Gateway is online, and you can extend it with skills, plugins, and code. It stores the preferences and project context you choose to save. The more deliberately you maintain that context, the more useful it can become.

We'll break each of these capabilities down in the next section.

How is a Claw different from Claude or ChatGPT?

You already use Claude, ChatGPT, and Claude Code. So what's new?

It lives in your messages. You don't open a separate app or go to a website. Your Claw shows up in WhatsApp, Telegram, Discord, or SMS—wherever you already talk to people. You interact with it the same way you'd text a friend: casually, frequently, in the flow of your day. There's no "AI time"—it's just part of how you communicate.

It can be extended. When you ask your Claw to do something it can't yet do, it can help you add a skill, plugin, or custom code. An outside service may still require an account, credentials, permissions, a public webhook, or manual configuration. Treat each new capability as software you need to inspect and test.

It operates proactively. Most AI tools wait for you to ask a question. Your Claw can act on its own. It can check your calendar each morning and tell you when to leave for work. It can watch your inbox and flag something urgent before you've even opened it. You set the rules for when it should reach out and when it should stay quiet—and it just runs, in the background, like a team member who knows their job. Those rules grant standing authority, so decide which sources it may read, which actions require approval, and where it should report what happened.

It has a personality you shape. Your Claw has a name and a personality. It is not a generic assistant with memory files. No two Claws are alike, because no two people are. The effect comes from workspace and memory files you build and maintain, rather than automatic improvement.

What does it have access to?

By default, your Claw only knows what you tell it in conversation. It doesn't have access to external accounts such as email or calendars until you explicitly connect those tools. Its access to files, web browsing, and commands depends on where its Gateway runs and which tools and permissions you enable.

When you're ready, you can ask it to connect to apps like Gmail, Google Calendar, Apple Notes, Notion, Todoist, and even your smart home. Ask what the connection requires, then verify the setup yourself. Depending on the service, you may need an account, credentials, permissions, or a separate tool. After connecting it, test the smallest possible read and inspect the access it received.

It can also make and receive phone calls through OpenClaw's voice-call plugin. You'll need a provider account, credentials, a phone number, a reachable webhook, and a test call. Require confirmation before the agent books or buys anything or shares personal information.

What are its boundaries?

Your Claw is powerful, but it's not omniscient. Here are its limits:

• •It can be confidently wrong, especially with specific facts, numbers, and recent events. Use it for drafts, research, and organization, and verify anything high-stakes.
• •It only knows what you've given it. If you haven't connected your calendar, it can't check your schedule. It will guess about your preferences if you have not specified something.
• •It's a collaborator, not an oracle. The best results come from back-and-forth conversation.
• •If you give it standing instructions, scheduled jobs, or broad tool permissions, it can act without a fresh request. You set the boundaries for what it can and can't do.
• •It requires upkeep. Models change, credentials expire, integrations break, schedules fail, and memory can become stale or contradictory.

What happened to Plus One?

Every originally built Plus One as a hosted personal agent. Our internal rollout showed that making one agent for every employee didn't remove the maintenance burden; it multiplied it. We are rebuilding Plus One around shared agents with defined jobs. You can read what we learned.

One of my (Dan's) favorite books growing up was Philip Pullman's The Golden Compass. In it, every character has a constant animal companion called a daemon that knows them intimately.

That's what a Claw feels like. From your very first interaction, you feel less like you're talking to a chatbot and more like you're dealing with something alive, something that you can have a real relationship with, something that gives you so much that you find yourself wanting to return the favor. It feels like yours. Working with them is different from working with tools such as Claude or ChatGPT in a few ways.

They become a mirror of the human who runs them

Claws quickly become a mirror of the humans who run them: Because you're using it every day, your Claw builds up context about your needs and personality. If you're a writer, like I (Dan) am, its perspective will be more literary and so will its suggestions. If you're a security-conscious executive or a programmer who appreciates brevity, your Claw will operate that way, too. The continuity depends on what you deliberately save, correct, and maintain.

The appeal is personal rather than universal: One agent can carry one person's context across projects, fit around idiosyncratic workflows, and give its owner control over the setup. For someone willing to maintain it, that can be worth the effort.

They can be extended

OpenClaw can help you assemble skills, integrations, and scheduled jobs for workflows that may be too specific for an off-the-shelf assistant. You could ask, "Message me once a day with a summary of the latest post on Every, a rundown of large-cap tech stock movements from the day before, and a quote from Annie Dillard." You will still need to connect the sources, inspect the permissions, test the output, and repair the workflow when something changes.

Teams may prefer a shared Claw

Organizations benefit from a different kind of continuity: a common operating method rather than one employee's private context. A shared agent can give a group access to the same skills, source connections, checklists, formats, and approval rules. When the maintainer fixes an integration or improves a workflow, everyone who uses the agent inherits the change.

OpenClaw describes this organizational pattern as a delegate: an agent with its own identity and credentials that acts on behalf of one or more people under explicit permissions. Give it a defined job and a named maintainer. Start with read-only access and drafts, then expand its authority only when the work requires it.

Ready to try it? There are several ways to get a Claw, depending on how hands-on you want to be.

Use the Mac app: Download the latest signed build from OpenClaw's releases. The app can install the matching command-line runtime, start a local Gateway, run model checks, and prompt for macOS permissions.

Use Windows Hub: On Windows 10 or 11, the native Windows Hub app is the easiest desktop path. It can provision an app-owned Windows Subsystem for Linux Gateway without changing your existing Linux distribution.

Run it on your laptop: If you have a Mac or Linux machine, type this one command in your terminal:

It installs OpenClaw locally and walks you through model setup and optional channel pairing. Then run openclaw onboard --install-daemon, check openclaw gateway status, and open openclaw dashboard. Get one clean reply in the Control UI before connecting a messaging channel. If you're on a Mac and want to keep it sandboxed, as in a computer on your computer, there's also a local VM option.

Run it on a server: If you want your claw running 24/7 without keeping your laptop open, deploy to a cloud provider like Fly.io, Hetzner, or Google Cloud. The hosting guide has step-by-step instructions for each. A server gives you uptime, not hands-off operation: Someone still needs to update the Gateway, protect credentials, and restore service after failures.

Whichever path you choose, the experience is the same: Your Claw shows up in your messaging app, and you start talking to it. Keep pairing enabled, approve only yourself, and add one channel before you add more tools.

In order to get the best use out of your Claw, you need to rewire your brain to think differently about getting things done. Here are the important factors in that mindset shift:

Think ‘delegation,’ not ‘search’

The biggest mistake people make is treating their Claw like a search engine. They type a question, get an answer, and move on, not taking full advantage of what the tool can do.

Think about it like you just hired someone. What would you delegate to a new team member? You wouldn't only ask them questions—you'd give them projects, explain your preferences, check their work, and work together to find a solution. The same skills that make someone a good manager make them good at using their Claw.

Start with what annoys you

Start by thinking about the most annoying part of your day. The thing you always forget. The task you keep putting off. The information you're always hunting for across three different apps.

That's your first project—as long as it happens often enough to repay the setup and upkeep.

It's a conversation, not a command

The people who get the best results don't fire off a single perfect prompt. They have back-and-forth conversations. They tell their Claw, "That's close, but make it shorter," or "Actually, I changed my mind about the format." Every exchange gives your Claw more context about what you want.

Think of your interactions with your Claw as if you are working with a colleague who's eager to help but needs to understand your taste and preferences first.

Expect the first attempt to be mediocre

Your Claw's first try at anything will rarely be exactly what you want, so don't make that an excuse for giving up early. Getting to the result you want requires iteration. A few back-and-forth messages get you where you want to be, which is still way faster than doing it yourself from scratch.

Let go of: ‘I could just do it myself’

Yes, you could Google the weather yourself or check your own calendar. What you forget is that all of those small tasks together become a heavy mental load. Your Claw is there to relieve you of that burden.

Count the maintenance, too. If a workflow saves five minutes but regularly needs credentials repaired, instructions updated, or failed jobs investigated, doing it yourself may still be cheaper.

Your Claw is powerful. It can send messages, read files, run code, and browse the web on your behalf. That's the whole point—but it means a few basic precautions are worth taking before you hand it more responsibility.

OpenClaw assumes one trusted operator for each Gateway; it isn't designed to isolate mutually untrusted users from one another. If an agent serves a team, give it its own identity and credentials, an explicit list of people allowed to direct it, and a named maintainer.

Lock the front door

Your Claw comes with pairing mode on by default: Anyone who messages it gets a one-time code, and you decide whether to let them in. Leave it that way. The moment you switch to "open" mode, anyone who finds your bot can talk to it—and anything they say gets treated as an instruction with access to whatever tools you've enabled.

Think of it like your phone. You wouldn't leave it unlocked on a café table. Your Claw should know exactly who it's talking to, and that list of people should be short.

After setup—and after any meaningful permission or integration change—run openclaw security audit --deep and review the findings.

Pick a private channel

Not all messaging platforms are equally suited for an AI that acts on your behalf. Telegram is currently OpenClaw's fastest phone setup, but the right channel depends on the identity controls and permissions you need.

If you add your Claw to group chats, turn on "require mention" so it only responds when someone @-tags it directly. Otherwise it's listening to every message in the room and treating each one as a potential instruction from you.

Sandbox first, unlock later

Your Claw can access your file system, execute code, and browse the web—but it doesn't need to be able to do all of that on day one. Start with messaging-only tools. When you need something more powerful, enable just that one and check the effective permission mode before you use it. Prefer deny, allowlist, or ask over broad permission to run commands on the host computer. Sandboxing is opt-in and reduces the blast radius, but it works alongside tool policy, approvals, and host isolation—not in place of them.

It's the same principle you'd apply to a new hire: Give them access to what they need for the job, not the master key to the building.

Be choosy about skills

Skills are like browser extensions for your Claw: They expand what it can do, but each one runs with your Claw's permissions.

Stick to well-known skills from ClawHub, a public registry for skills for Claws. Read a skill before you enable it. Check its audit status, findings, requested permissions and credentials, provenance, and version. A passing audit is a useful signal, not a guarantee. And if someone in a Discord server tells you to install their custom skill that "just works," treat that with the same skepticism you'd give a Chrome extension from an unknown developer.

Don't skimp on the model

The AI model your Claw runs on directly affects its security. Newer, more capable models are often better at recognizing prompt injection—when someone tries to trick your bot into ignoring its instructions and doing something it shouldn't. If your claw handles real tasks with real tools, use the best model you can. It's one layer of protection, not a substitute for narrow permissions and human approval.

Dan Shipper is the cofounder and CEO of Every, where he writes the Chain of Thought column and hosts the podcast AI & I. You can follow him on X at @danshipper and on LinkedIn.

Willie Williams is the head of platform at Every. You can follow him on X at @bigwilliestyle.