From zero to your own AI assistant. Free. No fluff.
Get Started →Claude Opus 4.7 released today. Identical pricing to 4.6 (£5/M input, £25/M output) but 70% on CursorBench. Direct OpenRouter access — one API key, switch via config. Set as your primary and let coding agents run it. Everything else stays on Kimi K2.5 ($0.50/M output).
Auth path: OpenRouter for everything (OpenAI-codex OAuth works with ChatGPT Plus for GPT-5.4 workloads). See the updated Model Guide.
Not a chatbot. Not a search engine. An AI agent that takes action. You give it goals, it delivers results, it improves. Like having your own company working for you — on any computer you already have.
Credit: Alex Finn — the best unbiased OpenClaw breakdown on YouTube
This isn't a search engine. You give it goals, it takes action, delivers results, and improves. It's a self-improving employee, not a glorified Google.
Telegram, WhatsApp, Discord — text your agent like texting a friend. Say "go build me this app" and it starts building on your computer.
It has agency. It decides what to do, monitors the web, acts without being asked. Like having your own company working for you.
Not a walled garden. Use any AI model, any tool, customise everything. It's completely open source.
📨 Get weekly OpenClaw tips & workflows
New features, model updates, and tricks — straight to your inbox. Free.
Answer 4 quick questions and get a personalised install plan.
This determines your install commands.
This determines which AI model we recommend.
Select all that apply.
Helps us tailor your setup recommendations.
Everything you need, in order. Click to expand each step.
OpenClaw runs on Node.js (version 18 or higher). Pick your operating system below for the correct steps.
Install Homebrew (if you don't have it):
/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"Then install Node:
brew install nodeVerify it worked:
node --versionbrew, install Homebrew first: brew.shDebian/Ubuntu: Use the official NodeSource repo to get a current version (default apt packages are often too old).
curl -fsSL https://deb.nodesource.com/setup_22.x | sudo -E bash -sudo apt-get install -y nodejsVerify it worked:
node --versionFedora/RHEL:
sudo dnf install -y nodejs💡 Strong recommendation: Install WSL first. Native Windows works but is officially "trickier" — stale plugins, missing exports, path issues. WSL (Windows Subsystem for Linux) gives you a proper Linux terminal inside Windows. Takes 10 minutes. OpenClaw runs perfectly on it.
Step 1 — Open PowerShell as Administrator (right-click PowerShell → "Run as administrator"), then run:
wsl --installThis installs Ubuntu inside Windows. Reboot when it tells you to.
Step 2 — After reboot, open the Ubuntu app (Start menu → type Ubuntu). Set a username + password when prompted. Password won't show as you type — that's normal.
Step 3 — Inside Ubuntu, install Node.js:
curl -fsSL https://deb.nodesource.com/setup_22.x | sudo -E bash -sudo apt-get install -y nodejsVerify:
node --versionFrom now on, every OpenClaw command in this guide goes into the Ubuntu terminal, not PowerShell.
Download the Node.js installer from nodejs.org and run it. Use the LTS version. Expect occasional plugin issues — if things break, switch to WSL.
For any other OS (BSD, Arch, Alpine, NixOS, etc.): install Node.js v18+ via your package manager, then proceed to the next step. If you're on one of these, you probably know what you're doing.
Official Node.js downloads: nodejs.org
One command. That's it.
npm install -g openclawsudo npm install -g openclawInitialise your workspace and configuration:
openclaw setupThis creates your workspace directory and configuration files. Follow the prompts. You can also run openclaw onboard for a more interactive guided setup.
Recommended: OpenRouter. One API key gives you access to every model — Claude, GPT, Gemini, Kimi, open-source, everything. Pay per token. Switch models anytime. This is the best way to run OpenClaw.
1. Go to openrouter.ai and create an account
2. Go to Keys → Create Key
3. Add credit ($10–20 to start — that'll last you a while if you optimise)
4. Run openclaw config → select your machine → model → OpenRouter → paste your API key
5. Set your default model. I recommend: Kimi K2.5 as default (cheap, capable), with Claude Opus 4.7 as fallback for complex tasks.
Why OpenRouter over direct API keys?
⚠️ What about Claude OAuth / subscriptions?
Anthropic removed OAuth access for third-party tools in April 2026. You can no longer use a Claude Pro/Max subscription to power OpenClaw. That flat-rate era is over. OpenRouter is the move now.
Free options: OpenRouter free models or Google Gemini free tier. Limited but enough to test if OpenClaw is for you.
Telegram is the easiest channel — connect this first, then set up everything else from within the conversation.
1. Open Telegram and search for @BotFather
2. Send /newbot and follow the prompts
3. Copy the bot token BotFather gives you
4. Paste that token into your OpenClaw agent — it'll configure itself automatically
5. Message your new bot in Telegram. You're live.
WhatsApp uses a bridge connection. Setup involves scanning a QR code from your terminal.
Check the official docs for the latest WhatsApp setup guide — it updates frequently.
1. Go to Discord Developer Portal
2. Create a new application and bot
3. Copy the bot token
4. Invite the bot to your server with the generated OAuth2 URL
5. Add the token to your OpenClaw config
Start the gateway:
openclaw gateway startOpen Telegram and message your bot. It should respond within seconds. You're live.
Once Telegram is working, you can set up other connections (WhatsApp, Discord) by simply asking your agent to help you connect them. It'll walk you through it.
If it doesn't respond — check the terminal for error messages. Most common issue: API key or OAuth not configured correctly.
OpenClaw uses markdown files to define your AI's personality and behaviour:
SOUL.md — Who your AI is. Its tone, personality, how it talks. Write this like you're describing a person.
AGENTS.md — Operating rules. How it should handle memory, group chats, proactive behaviour.
USER.md — Who you are. Context about you so the AI can be more helpful.
These files live in your workspace: ~/.openclaw/workspace/
🧠 Pro tip: Brain dump then reverse prompt. Dump everything about you — goals, ambitions, interests, daily workflows, career info — into your agent. Then ask: "Based on what you know about me, what are 10 workflows and automations you can implement for me right now?" This is the most powerful exercise in all of AI.
📝 Daily journals for memory: Have your agent create daily journal entries logging everything you discuss. This dramatically improves its memory over time. Just say: "Every day, create a journal entry and log anything important we discuss."
⚠️ Skills safety: Don't blindly install other people's skills from ClawHub. Instead — send your agent the link to a skill and say "Check this out and build your own version." That way you know it's safe. If your agent does something well, say "Turn that into a skill." If it does something poorly, say "Figure out a better way and write a skill for yourself."
You're set up. Now what? Do these three things to see the power immediately.
Before you start — pull all your personal context out of whatever AI you've been using. Paste this into ChatGPT, Claude, or whatever you're currently on:
I'm moving to a new AI system. Please compile everything you know about me into a structured document: my name, goals, personality traits, communication style, career details, interests, daily routines, preferences, projects I'm working on, and anything else you've learned about me. Format it as a clean profile I can paste into a new system.Have your agent send you a summary every morning — news, tasks, whatever matters to you. This runs automatically using a cron job.
Set up a cron job to send me a daily brief every day at 8am. Include the top 3 news stories in [your industry], my upcoming tasks for the day, and anything I should know about. Schedule this as a recurring automated task.Ask your agent to build you a personal dashboard — a command centre for your life and work. It vibe-codes the whole thing for you.
Build me a mission control dashboard. A local web app where I can see my tasks, calendar, notes, and any tools we build together. Make it look clean and modern. Dark mode. I want to be able to add new tools to it over time — like a task board, journal, research tools, whatever I need.Dump everything about yourself — goals, interests, workflows, career — then ask what it can do for you.
Here's everything about me: [paste your goals, what you do, your daily routine, what you're building, what frustrates you]. Based on what you now know about me, what are 10 workflows and automations you can set up for me right now?The biggest upgrade you can make to your OpenClaw setup. Turn it from a chatbot with amnesia into an agent that remembers everything — clients, projects, preferences, mistakes, patterns. This is what separates toy AI from a real operating partner.
🧠 The problem you're about to solve: OpenClaw's memory out of the box is limited. It forgets between compactions. It loses context across sessions. Small tasks get dropped. You've probably already felt it.
The fix: An Obsidian vault as the agent's long-term memory. Always accessible, selectively loaded, survives compactions, shared across agents. This is how I run my entire operation.
Credit: Alex Finn — the cleanest explanation of this system on YouTube.
OpenClaw ships with layers 1, 2, and 4. You're adding layer 3 — the vault. This is what makes everything click.
Sticky-note level. Your agent's name, important file paths, core identity. Injected into every prompt. Keep it tiny.
AGENTS.md = the rules file. SOUL.md = personality and tone. Both loaded every session. This is how your agent knows how to act.
Full second brain. Daily logs, mistakes file, project notes, people files, decisions. Selectively loaded — not injected every prompt. Agent pulls from it on demand and checks it after every compaction.
Built-in indexed history of every session and cron run. Semantic search finds exactly the context you need without loading everything.
One markdown file per day (memory/YYYY-MM-DD.md). Agent writes everything important from each session automatically. This alone fixes 80% of your memory issues.
Every time you correct the agent ("you did that wrong"), it logs it to mistakes.md. Now it doesn't repeat the same errors. Compound learning.
One file per person. Clients, partners, family. Relationship history, preferences, past interactions. Your agent gets proper context next time their name comes up.
One file per project. Status, next actions, blockers, decisions made. "Pick up the X project from last week" actually works now.
Market intel, book notes, breakthrough ideas, conversations worth preserving. Everything that used to vanish into chat history — captured and searchable.
If you run multiple agents (OpenClaw + Hermes, or multiple OpenClaw agents), this folder is the shared workspace. They all read + write to it. One agent picks up where another left off.
Obsidian is a free, local-first markdown editor. Download from obsidian.md. Install it like any app.
Create a new vault — name it something like AI-Brain. Note the full file path where you create it. You'll give this path to your agent in step 3.
KJ Rainey has a solid vault template that gets you productive immediately. Built for Claude but works perfectly with OpenClaw.
Download KJ Rainey's template →
Unzip it, drag into your Obsidian vault. Customise the folders to match your life — People, Projects, Knowledge, etc.
Copy this prompt and paste it into your OpenClaw agent. It will build the whole system into your AGENTS.md automatically.
I want to add a 4-layer memory system to my setup using an Obsidian vault. My vault is located at [REPLACE WITH YOUR VAULT PATH].
Set this up:
1. Read my current AGENTS.md and understand my existing rules.
2. Add a "Memory" section to AGENTS.md explaining the 4-layer system:
- Layer 1: Built-in memory (tiny, always-on facts)
- Layer 2: AGENTS.md + SOUL.md (rules + personality)
- Layer 3: Obsidian vault (selective, on-demand)
- Layer 4: Session search (built-in history)
3. Create the vault structure if it doesn't exist:
- memory/ (daily logs, YYYY-MM-DD.md format)
- People/
- Projects/
- Knowledge/
- Agent-Shared/
- mistakes.md
- working-context.md
4. Add rules so you automatically:
- Write a daily log at the end of each day summarising what we did
- Log corrections to mistakes.md whenever I tell you something was wrong
- Create/update People files when a new person is mentioned
- Create/update Project files when a new project starts or a milestone hits
- Update working-context.md at task start and task completion
- Check the vault after every compaction to recover context
- Pull from the vault on demand when I reference past work
5. Use wiki links like [[filename]] to connect files.
6. Write this entire system into AGENTS.md so it survives restarts and is followed automatically going forward.
Validate it worked by creating one sample daily log, one sample person file, and one sample project file.Validate it worked: after the agent finishes, check that:
AGENTS.md now has the memory rules inside it (not just discussed in chat)When OpenClaw compacts mid-conversation (it will), your agent reads the vault to recover context. You won't even notice compactions happened.
You can reference work from any past date and your agent will find it. The small projects that used to vanish are now retrievable.
Mistakes file means your agent stops repeating the same errors. Every correction makes it sharper. This is the compounding edge.
If you run two agents (mine is OpenClaw + Hermes), the Agent-Shared folder lets them hand off work. One drafts, the other reviews.
My vault has 217+ markdown files. Organised into: People (clients, partners, contacts), Projects (active + parked), Knowledge (market intel, book notes, learnings), Daily (one file per day going back months), Agent-Shared (collaboration space), plus a dedicated space per agent for its working context.
Auto-update rules baked into AGENTS.md: when I mention a new person, file gets created. New project, file gets created. Decision made, logged. Mistake surfaced, logged. No manual admin.
Cost: nothing extra. Obsidian is free, files are local, the only cost is the tiny amount of tokens used to read/write the vault.
Don't know what to say? Start with these. Copy, paste into Telegram, hit send.
This dramatically improves your agent's memory over time.
Every day, create a journal entry in your memory folder logging anything important we discuss. Include decisions, tasks, ideas, and anything I'd want to remember later.Once Telegram is working, just ask your agent to set up Discord.
I want to connect you to my Discord server. Walk me through the setup step by step.Your agent can spin up sub-agents to research multiple angles simultaneously, then compile everything into a single report.
I need deep research on [topic]. Spin up sub-agents to research this from multiple angles: market size, key players, recent developments, risks, and opportunities. Compile everything into a structured report with sources. Take your time and be thorough — I want this to be comprehensive.Your agent uses cron jobs to run tasks automatically on a schedule — no manual triggering needed.
Set up the following cron jobs for me: 1) Every Monday at 9am, send me a weekly priorities review. 2) Every Friday at 5pm, compile a summary of everything we accomplished this week. 3) Every day at 8pm, check my inbox and flag anything urgent. Schedule all of these as recurring automated tasks.Your agent can vibe-code full applications. Describe what you want and let it build.
Build me a [describe what you want — e.g. "client tracker", "habit tracker", "invoice generator"]. Make it a local web app with a clean dark UI. Include [specific features you want]. Store data locally. I want to be able to run it from my browser. Take your time and make it look professional.This is what makes it yours. Define who your agent is, how it speaks, and what it prioritises.
Update your SOUL.md with the following: Your name is [name]. You're my [role — e.g. chief of staff, strategic advisor, personal operating system]. Your tone is [describe — e.g. direct, sharp, no fluff, grounded, slightly witty when appropriate]. You always [key behaviours — e.g. challenge weak thinking, prioritise clarity over politeness, close loops]. You never [things to avoid — e.g. use corporate jargon, give generic advice, sugarcoat]. When I'm under pressure, you [how to respond — e.g. simplify, ground me, prioritise the next move].When your agent does something well, lock it in so it can do it again. When it does something poorly, make it learn.
That [thing you just did] was exactly what I needed. Turn that into a skill — write a SKILL.md file that captures the approach, the format, and the quality standard so you can reproduce this anytime I ask. If there's anything you'd improve, include that too.Which AI brain should you use? Here's the honest breakdown.
⚡ The best way to access models now: OpenRouter — one API key, access to every model. Claude, GPT, Gemini, Kimi, open-source models, everything. Pay per token. No subscriptions. Switch models any time. This is the move after Anthropic killed OAuth.
Anthropic · via OpenRouter
Best for: Getting shit done. 70% on CursorBench. The most capable model available. This is what I run daily. A mile ahead of everything else.
~$15/MTok input · ~$75/MTok output (via OpenRouter)
Moonshot AI · via OpenRouter
Best for: Budget powerhouse. Surprisingly capable for the price. Great as a default model — handles most tasks well. Set this as your primary and use Opus as a fallback for complex work.
~$1-2/MTok (via OpenRouter) — fraction of Claude's cost
Anthropic · via OpenRouter
Best for: The middle ground. Smarter than Kimi, cheaper than Opus. Great for most tasks.
~$3/MTok input · ~$15/MTok output (via OpenRouter)
OpenAI · via OpenRouter
Best for: Arguably the smartest model available. But — it gets stuck on complex tasks ~80% of the time. You can stub its toe and it rolls over. Claude crawls across the finish line with a broken leg.
Via OpenRouter
Ollama / LM Studio / Nvidia Neotron
Best for: The future. 100% private, free, no one can take it away. Currently needs $4K+ hardware (DGX Spark) for good performance. In ~6 months, $1-2K hardware will run Opus-level models locally.
€0 (hardware cost only)
Now that you're paying per token, here's how to keep costs sane.
Every message you send includes your entire conversation history + workspace files. If your MEMORY.md is 10,000 tokens, your SOUL.md is 3,000 tokens, and your conversation has 50 messages — that's a LOT of tokens being sent with every single message. The AI reads everything, every time.
Instead of loading every memory file in full, OpenClaw can search by meaning and pull only what's relevant. This is the single biggest cost saver.
Set up semantic memory search using Gemini embeddings. Add memorySearch config with provider gemini, hybrid search enabled, MMR diversity, and temporal decay of 30 days. This way you only pull relevant context instead of loading everything.Your workspace files (MEMORY.md, SOUL.md, USER.md, AGENTS.md, TOOLS.md, HEARTBEAT.md) are injected into every message. Keep them focused:
Set a cheap model as default and expensive models as fallbacks. Your agent uses Kimi K2.5 for routine stuff and only hits Opus when you specifically ask or the task demands it.
Set my default model to openrouter/moonshotai/kimi-k2.5 with fallbacks to openrouter/anthropic/claude-sonnet-4.6 and openrouter/anthropic/claude-opus-4.7. This way I get cheap responses for simple tasks and can switch to Opus when I need the big guns./status regularly to monitor token usage per sessionReal talk. No fear-mongering, just what you need to know.
Your API keys give access to paid AI services. Never share them, never commit them to Git, never screenshot them in public.
OpenClaw runs on your machine. Your conversations go directly to your AI provider — not through a middleman's server.
Messages → your machine → AI provider (Anthropic/OpenAI/Google). That's it. No OpenClaw server in between.
OpenClaw can run commands on your machine. Review what tools you enable. Start conservative, expand as you get comfortable.
Local: Your laptop. Easy but turns off when you close it. VPS: Always-on server. Runs 24/7 but needs basic server security (SSH keys, firewall).
This is the #1 thing. Personal accountability. OpenClaw only does what you ask it to do. It won't randomly leak your photos on Facebook or tweet your passwords — unless you tell it something that implies it should. Before every prompt, think: could this cause something I don't want? That's literally the whole security model. The vast majority of the millions of people using OpenClaw have never had a single security issue. Be mindful. That's it.
I run OpenClaw on my local Mac with Claude Opus 4.7 via OpenRouter. Telegram as primary channel. Conservative tool permissions. Semantic memory search enabled to keep token costs down. Kimi K2.5 as default model, Opus 4.7 for complex tasks.
📬 Stay in the loop
Model updates, new workflows, and things worth knowing — straight to your inbox. No noise.
Everything else you might need.
Don't panic. OpenClaw is just a folder on your computer with markdown files. Here's how to fix 100% of issues:
Open Claude Code or Codex in your OpenClaw folder and say:
My OpenClaw isn't working. Please go in and see why.The AI agent goes through your config files and finds the issue. 99% of the time it's a config issue from someone editing files manually. This fixes it every time.
Found something useful? Something missing? Let me know — anonymous or not, your call. Nothing gets posted publicly.