OpenClaw for those new here

OpenClaw can feel overwhelming for anyone landing here for the first time. The fieldguide quickly assumes you know what a gateway is, what a session is, and why you would want an agent that does not live in a browser tab. What follows is a calmer entry: not oversimplified beneath the truth, but explanation without assumption. For anyone with no prior knowledge of AI agents, self hosting or messaging runtimes, who wants to understand it at their own pace.

01 · WHAT IS AN AI AGENT, REALLYThe difference between a chatbot and an agent

Everyone has talked to ChatGPT or Claude.ai by now. You type a question, you get an answer, you close the tab. Next time you open it, the conversation starts over. That is a chatbot: something that responds to what you send and then sits idle.

An agent is something different. An agent has access to files, to the internet, to specific tools. It can do things, not only talk about them. “Send Tom an email saying the meeting has been moved” is, for a chatbot, a request to compose the email text you then copy. For an agent, it can be an instruction it executes itself, provided it has access to your email account.

The second difference: an agent remembers things between conversations. A chatbot forgets at each new session. An agent knows what you asked last week, knows your preferences, knows which projects you are working on. That memory is what makes it feel like a digital colleague, not a tool.

The third difference: an agent can be proactive. Not only respond, but do something on its own. Send a briefing every morning, write a weekly overview every Monday, give a reminder when you risk forgetting something.

An AI agent then is: an AI with tools, memory and its own time. Not entirely like a human, but in the neighbourhood.

02 · WHY YOUR OWN RUNTIME, NOT CHATGPTWho OpenClaw is for, and when it is not

ChatGPT and Claude.ai are fantastic AI tools. For most people they are enough. So why want something more complicated?

Three reasons.

Your data stays with you. OpenClaw runs on your machine. Your conversations, your files, your memory live on your disk. The only data that leaves the machine is the short piece going to the LLM model for the thinking. For people working with confidential information (legal, medical, financial, personal) that is a real argument.

You do not live in a separate tab. OpenClaw sits inside the messaging apps you are already in: WhatsApp, Telegram, Discord, Signal, Slack. You do not need a separate app to talk to your AI; you message it the way you message a colleague. For anyone switching between tools all day, that saves a lot of context loss.

You stay in control. Which model to use (Claude, GPT, Gemini, local Llama), which skills to activate, which channels to connect, how memory works; all your choice. With ChatGPT and Claude.ai you are stuck with what the vendor allows.

For whom is this not the right choice? For anyone with occasional questions and no need for continuity, ChatGPT is fine. For anyone who does not want to run Docker or configure anything, also fine. OpenClaw requires some comfort with the command line and with technical choices. It is not for everyone, and that is not a failure.

03 · THE ESSENCE IN FIVE SENTENCESWhat OpenClaw fundamentally is

For anyone without time for the whole fieldguide, here is OpenClaw in five sentences:

1. OpenClaw is a local runtime for personal AI agents, which you control via your regular messaging apps.
2. An agent is a claw with its own personality, its own memory, and access to tools it uses on your behalf.
3. All data lives on your machine; only the conversations with the LLM model leave your network.
4. The agent can work proactively through heartbeats and cron, not only reactively to your questions.
5. OpenClaw is built to grow with you: more agents, more skills, more channels, without the architecture having to change.

Anyone who carries these five sentences can reduce every page of the fieldguide to something they already know.

04 · THE THREE THINGS YOU NEED TO GRASP NOWThe minimum required mental models

OpenClaw has eleven primary abstractions. To start you need to know three: the other eight come along when you bump into them.

A claw is an individual agent. With its own name, personality, memory and channel connections. One OpenClaw installation can run multiple claws at once (one for work, one for home), but you start with one. Picture it as an assistant who knows your stuff.

A channel is a connection to a messaging platform. Telegram is a channel, WhatsApp is a channel, Slack is a channel. One claw can be present on multiple channels (like a person reachable on multiple phone numbers). The claw knows which channel a message arrives on and sends its reply back the same way.

A skill is a reusable ability. A skill is a folder with a markdown file in it describing what the skill does and how. “Send an email” can be a skill. “Read my calendar” can be a skill. “Summarise a PDF” can be a skill. You activate skills, and from then on the claw knows when it can use them.

With these three you have the minimal mental map: a claw (the agent) lives in your channels (the communication) and can execute skills (the actions). Everything that follows in the fieldguide is built on these three.

05 · READING THE FIELDGUIDEA recommended order

The official fieldguide follows a logical order for anyone already installing OpenClaw. For anyone who first wants to understand, a different route is nicer. A recommended reading order for the first time:

Read first (understanding where you are heading):

• What OpenClaw actually is (introduction, ~10 min)
• What makes a Claw unique (the four core values: personality, memory, proactive, self extending)
• Architecture in brief (the three layers, no more than the overview)
• Concepts you will encounter (glossary to keep handy)

Read only when you are about to install:

• Part 1: Installation and first start
• Part 2: Security (read before going further, not after)
• Part 3: Connecting your first channel (start with Telegram, not with WhatsApp)
• Part 4: LLM providers (pick one to start with)

Read when you want to go deeper:

• Part 5: Personalisation (writing SOUL.md, USER.md, MEMORY.md)
• Part 7: Skills (which to install, which later)
• Part 6: Memory (how it works, what the three layers are)

For later (you really can skip until you need them):

• Part 8: Multiple channels at once
• Part 9: Multi agent
• Part 10: Scheduled tasks with cron
• Part 11: Browser automation
• Part 13: Docker sandboxing
• Part 14: Hooks and automation
• Part 15: Native apps

The rest (CLI reference, configuration overview, troubleshooting, updating) is reference: needed when you have to look something specific up, not for reading through.

06 · WHAT YOU DO NOT YET NEED TO GRASPParts that can wait

Some parts of OpenClaw are for anyone going beyond the first agent. You do not need to understand them early.

Hooks. Hooks are handlers that respond to events (a message arriving, a session started, an email received). They are powerful but conceptually heavy. For your first setup: ignore.

Multi agent and routing rules. Multiple claws on one installation, each with their own workspace and channels. Only relevant once you want more than one clearly bounded context. Start with one.

Custom skill development. Writing your own skills is fun but not necessary. The built in skills and those from ClawHub cover most of it.

Docker sandboxing details. The default non-main mode is fine for personal use. Only relevant when you want a public bot or experiment with unknown skills.

Native app integration and device capabilities. The iOS and macOS apps offer camera, location, voice mode, Apple Watch pairing. All optional. For starting out you do not need to know about them.

Memory wiki, active memory, dreaming, REM backfill. Advanced memory systems. For starting out the three memory layers are enough (MEMORY.md, daily notes, vector index).

That sounds like a lot, but the point is: OpenClaw is an extensive system of which a small part is enough for most users. Build out where you run into limits, not just to be safe.

07 · FIRST STEPS AFTER READING THISConcretely what you can do now

Three suggestions for what you could concretely do next, in increasing commitment.

Option 1: First read through on the Placebo page. The Placebo Pill shows the complete fieldguide unchanged. Read the first three sections (What OpenClaw is, What makes a Claw unique, Architecture in brief) to get a fuller picture before you install anything.

Option 2: First look through another lens. The Mental Pill gives a conceptual blueprint; good for anyone who wants the thinking model sharp before the technology starts. The Security Pill helps you understand the risk layer before you connect anything.

Option 3: Start installing. Follow the Build Pill, which presents the installation as a staged hands on path. For anyone who learns by doing: this is the direct route. Allow about two hours for the first setup to a working first agent.

None of these three is wrong. The right choice depends on how you learn: read first, understand first or do first. OpenClaw itself forgives all orderings.