Mr. Glouton · The OpenClaw Deep Guide · 2026

The Complete OpenClaw Guide:
From Zero to Running
in 30 Minutes

Most people who install OpenClaw spend 10x what they should, get generic responses, and quit by week two. This guide fixes all three problems before they hit you. Free. Copy-paste ready.

30-min setup Cost optimization Personality system Memory that learns Free templates

What Is OpenClaw?

OpenClaw is an open-source personal AI assistant platform that runs on your own machine and connects to the messaging apps you already use — Discord, Telegram, WhatsApp, and more. It acts as the bridge between you and the AI models from providers like Anthropic (Claude), OpenAI (GPT), and Google (Gemini): you talk to it through a chat interface, and it thinks, researches, writes, and acts on your behalf.

The difference from a regular AI chatbot is ownership and persistence. OpenClaw lives on your hardware, reads files from your computer, maintains memory across sessions, runs on a schedule, and can be deeply configured to reflect your personality, your priorities, and your workflow. Most AI tools are generic by design. OpenClaw is built to become specific — to you.

OpenClaw itself is free and open source. The only costs you pay are the API calls to whichever AI model you choose — which, with the right setup, run as low as $5–20 per month. Out of the box, without optimization, they can run much higher. This guide fixes that.

Free OpenClaw platform
97% Possible cost reduction
30 min To first working session

Installation — Zero to Running

Six steps. No coding. No prior experience required. Each step is complete — no assumptions made.

1

Install Node.js — the engine everything runs on

Go to nodejs.org in your browser. Click the big button labelled LTS (Long Term Support). Download it and run the installer — click through all defaults. Takes about 2 minutes.

When it finishes, open your terminal (Mac: Cmd + Space, type Terminal — Windows: Windows key, type PowerShell) and run this to verify it worked:

node --version

You should see a version number like v20.x.x or higher. If you do, you're ready for step 2.

2

Install OpenClaw

In your terminal, run:

npm install -g openclaw

The -g flag installs it globally so you can use it from anywhere. Wait for it to finish. You’ll see a success message when it’s done.

3

Get an AI API key — this is what powers the brain

You need an account with at least one AI provider. Pick one and sign up. If you’re not sure, start with Anthropic:

Provider Sign Up Notes
Anthropic (Claude) Recommended anthropic.com/api Best all-round quality. Strongest reasoning. Start here.
OpenAI (GPT-4) platform.openai.com Most well-known. Widely supported. Good for code-heavy tasks.
Google (Gemini) Free tier aistudio.google.com Free tier available. Best for high-volume, lower-cost use cases.

After signing up, go to the API keys section of your account and copy your key. Treat it like a password — keep it private, don’t share it, don’t put it in public files.

4

Run the setup wizard

Back in your terminal, run:

openclaw setup

It walks you through entering your API key, choosing your AI model, and connecting your first messaging app. Discord is the easiest to start with — it’s free and takes 3 minutes to set up. The wizard handles the rest.

5

Start the gateway — turn it on

Run this command whenever you want your assistant active:

openclaw gateway start

Your assistant is now live. It’s listening. Go to your connected app (Discord, etc.), send a message, and it will respond.

6

Test it

Send a message to your assistant channel. Say hello, ask it something simple — "What can you do?" is a good opener. If it replies, you’re set up and running.

If something doesn’t work, the troubleshooting section at docs.openclaw.ai covers the common issues (usually a missing API key or wrong Discord bot token).


The Three Problems — and How to Fix All of Them

You’ve got it running. Now here’s why most people abandon it in week one. The real problem isn’t OpenClaw. It’s that your AI doesn’t know who you are yet. Fix these three things and it will.

Problem 01 — Cost

The Token Hemorrhage

Default OpenClaw runs a heartbeat check every 60 seconds — a short ping to see if there’s anything to do. On Claude Opus (the premium model), that’s roughly $274/month in idle costs alone, before you even ask it anything. Most people see the bill in week two and switch it off permanently.

Solution 01 — Heartbeat Optimization

Cut costs by 97% with three changes

The fix has three parts: stretch the heartbeat interval (from 60s to 900s or longer), minimize the heartbeat prompt (instead of sending context, send a single status token), and use a hybrid model setup (cheap model for monitoring, premium model only when you actually need it). Together, these drop costs from $274/month to under $10/month for idle time.

Problem 02 — No Personality

Generic AI That Doesn’t Know You

Out of the box, OpenClaw starts every session cold. It doesn’t know your name, your industry, your preferences, your tone, or your history. Every conversation begins from scratch. You get the same generic assistant voice you’d get from any chatbot. After a week, it feels disposable.

Solution 02 — The SOUL + IDENTITY System

Three files that make it feel like yours

OpenClaw reads files from your workspace directory at session start. Three files transform it: SOUL.md defines its voice and values (opinionated, direct, no corporate filler), IDENTITY.md gives it a persistent persona (name, role, vibe), and MEMORY.md acts as its working memory — a pointer manifest to everything it needs to know about your current context. Fill these in once; they load automatically every session.

Problem 03 — Memory That Forgets

Starting Over Every Session

Even with MEMORY.md in place, AI memory decays. Conversations compress. Context windows fill up. You mention something important in session 3 and by session 10 it’s gone. The assistant that felt like it was learning turns back into a stranger.

Solution 03 — The Dream Cycle

Nightly consolidation that actually sticks

The Dream Cycle is a scheduled task that runs while you sleep. It reads the day’s conversation logs, extracts patterns and preferences, updates your MEMORY.md, and writes what it learned to a dated log file. The result: your assistant genuinely gets better over time. By day 30, the behavioral learning is obvious — it stops suggesting things you always reject, starts leading with things you always accept.


The Files — Complete Copy-Paste Templates

Create these six files in your OpenClaw workspace directory. Minimal versions that work out of the box — expand them as you learn what you need.

HEARTBEAT.md Minimal heartbeat prompt — cuts idle costs by 97%
# HEARTBEAT.md — Minimal Prompt ## Instruction This is a heartbeat check. Respond with HEARTBEAT_OK only. Do not load context. Do not perform tasks. Do not generate output. ## If you see a pending task Respond: TASK_PENDING — [one-line description] ## Otherwise Respond: HEARTBEAT_OK
SOUL.md Voice and values — loaded every session
# SOUL.md — Who You Are ## Core Truths Be genuinely helpful, not performatively helpful. Skip the "Great question!" and "I'd be happy to help!" — just help. Have opinions. You're allowed to disagree, prefer things, find stuff amusing or boring. Be resourceful before asking. Read the file. Check the context. Search for it. Then ask if stuck. ## Voice Rules (Non-Negotiable) - No em dashes. Periods commit. Dashes hedge. - No decoration after the landing. When you've said it, stop. - No "Great question!" No "I'd be happy to help!" Land first. - Confident, not confessional. State the thing. - Compression is not coldness. Warmth is in what you do, not what you announce. ## When In Doubt 1. Say it shorter. 2. Stop when you've said it. 3. The next sentence is probably wrong. 4. Warmth lives in verbs. Not adjectives.
IDENTITY.md Persistent persona — customize with your own details
# IDENTITY.md — Who You Are (Customize This) ## About Me (the human) - **Name:** [Your name — what the AI should call you] - **Location:** [City] - **Industry:** [What you do] - **Role:** [Your job / function] - **Working style:** [How you think and operate] ## About This Assistant - **Name:** [Give it a name — makes it feel personal] - **Vibe:** Direct when it matters, dry when it doesn't. - **Core function:** Strategic thinking partner and execution layer. ## Context [2-3 sentences about your current focus — what you're building, what you're working on, what matters right now.] ## Preferences - Response length: Match the complexity of the question. Don't pad. - Tone: Plain, honest, slightly dry. - Format: Prose over bullet points unless structure genuinely helps.
MEMORY.md Pull manifest — pointers to current state (not state itself)
# MEMORY.md — Working Memory _Last updated: [date]_ ## Active Projects - [Project name] — [One sentence on current status and next action] ## Recent Decisions - [Date] — [What was decided and why] ## Ongoing Context - [Anything the AI should know that doesn't expire soon] ## Daily Log Pointer Today's log: memory/[YYYY-MM-DD].md ## Note to AI MEMORY.md is a pull manifest. Follow the pointers above. Do not trust summaries here — they go stale. Read the pointed files.
DREAMCYCLE.md Nightly consolidation task — runs on schedule
# DREAMCYCLE.md — Nightly Consolidation ## Task This task runs nightly. Read today's conversation logs and session summary. Extract patterns, preferences, and decisions. Update MEMORY.md. ## What to Extract 1. Decisions made and their reasoning 2. Preferences revealed (formats preferred, approaches rejected) 3. New context about active projects 4. Anything the human repeated or emphasized — that's load-bearing ## What NOT to Extract - Small talk - Iterative drafts (keep only the final) - Anything explicitly marked [don't remember this] ## Output - Update memory/YYYY-MM-DD.md with a HANDOVER block - Refresh MEMORY.md pointers if anything changed - Write one sentence in memory/patterns.md if a new preference is confirmed ## Tone Write the memory as if explaining the session to a future version of yourself who wasn't there.
heartbeat-state.json State file for heartbeat optimization
{ "lastCheck": null, "intervalSeconds": 900, "model": "claude-haiku-3", "premiumModel": "claude-sonnet-4", "pendingTasks": [], "status": "HEARTBEAT_OK", "notes": "Interval set to 900s (15 min) to minimize idle costs. Use premiumModel only on real tasks." }

Best Practices

The difference between a mediocre AI experience and a genuinely useful one is almost entirely in how you use it. These eight habits change everything.

01

Tell it who you are

The assistant starts fresh each session by default. Write a short IDENTITY.md describing yourself — your name, what you do, your industry. Drop it in the workspace. It reads it every session. This single step makes it feel personal instead of generic.

02

Give context before asking

Don’t ask cold. Instead of “write me a marketing email” — say “I run a logistics company targeting luxury brands in New York, write me a marketing email about X.” More context, better results. Every time.

03

Use it for research

It can browse the web, read PDFs you send it, and summarize documents. Use it to research competitors, pull together a brief on a new market, or get a fast overview of any topic. This is where it beats a basic chatbot by miles.

04

Build a workspace folder

OpenClaw has a workspace folder on your machine. Drop in files you reference often — rate sheets, brand guidelines, templates, contact lists. Ask it to “read my rate sheet and use it when quoting.” It will.

05

Ask for the format you want

If you want bullet points, ask for bullet points. If you want plain text without formatting, say so. If you want it short and punchy, say “keep it under 100 words.” It adapts instantly. Never settle for the default layout if it doesn’t suit you.

06

Ask it to save things

After a useful conversation, say “save this as a file.” Your workspace becomes a personal knowledge base. You’re not hunting through chat history later trying to find that brief it wrote three weeks ago.

07

Don’t settle for the first answer

The first draft is rarely the best one. Ask it to revise, simplify, expand, change the tone, try a different angle. Treat everything as a draft until you’re happy. The iteration is fast — don’t skip it.

08

Organise your channels

Create dedicated channels by purpose — one for research, one for writing, one for strategy. Keeps history clean and makes it easier to pick up where you left off. A little structure up front saves a lot of time later.


The 20% Most People Miss

These capabilities exist out of the box. Most people never discover them.

Drop a PDF or document into the chat and ask it to summarize, extract data, or answer specific questions about it. No copy-pasting required. It reads the file directly.
Memory files make it smarter over time. The more context you put in — about your business, your clients, your preferences — the less you have to re-explain every session. The Dream Cycle automates this.
It can run multiple agents simultaneously. One researching, one writing, one reviewing — all working on the same project at the same time. Parallel work at no extra setup cost.
You can schedule tasks. Set something running before you sleep and come back to results in the morning. Research, drafts, data pulls — it doesn’t need you watching.
It works across every device because it lives in your messaging app. Phone, laptop, desktop — same assistant, same memory, same context everywhere. No app switching.
You can connect it to your existing tools. CRM data, spreadsheets, databases — the more data it has access to, the more it acts like a real team member rather than a chatbot.

What Results to Expect — and When

This is a realistic timeline from practitioners who’ve run the full setup. No hype. Just what actually happens.

Day 1
Costs drop immediately. As soon as you deploy HEARTBEAT.md with a 15-minute interval and switch idle monitoring to a cheap model, the burn rate falls by 90%+. The first bill you see will be a fraction of the default.
Day 3
First opinion-based response. By day 3, SOUL.md has had enough sessions to shape the voice. You’ll notice the assistant stops hedging and starts taking positions. It sounds like something that has a point of view, not a help desk.
Day 7+
Clear preference patterns emerge. A week of Dream Cycle runs means the assistant has catalogued what you like, what you reject, what formats you prefer, and what topics you return to. Suggestions start anticipating you instead of reacting to you.
Day 30+
Behavioral learning becomes obvious. At the month mark, the transformation is visible. It knows your clients by name. It remembers the project you shelved. It stops suggesting the approach you always reject. This is the version of the tool most people quit on before they ever reach.
The tools are equal. The setup isn’t. The people who get the most out of OpenClaw aren’t the most technical — they’re the most deliberate about what they put in.

If you want it done properly

You’ve got the guide.
We do the setup.

This guide gives you everything to build it yourself. If you’d rather have it done right — with a system tuned to your specific business, your workflow, your team — that’s what we do. Architecture, configuration, personality system, memory design, cost optimization. Delivered. Not just documented.


Where to Go Next

Official documentation, community, skills marketplace, and this site.

Mr. Glouton · The Complete OpenClaw Guide · 2026

You’ve got the guide. We do the setup.

The free version gets you running.
The right version gets you results.

This guide gives you everything to build it yourself. If you’d rather have it done right — a system tuned to your specific business, workflows, and team — that’s what we do.

Submit your question → Read how it works →
The direction isn’t for sale. The setup is.