DEV Community

Jim L
Jim L

Posted on

I Set Up OpenClaw in 30 Minutes — Here's the Guide I Wish I Had

OpenClaw has 68K+ GitHub stars and a lot of hype. But most guides skip the parts that actually matter: how to make it not forget everything, how to stop it from picking the wrong tools, and how to keep it on track during long tasks.

I spent two weeks getting my setup right. This is what I learned.

What OpenClaw Is (30-Second Version)

OpenClaw is an open-source AI assistant that runs on your machine and actually does things — not just talks about them. Built by Peter Steinberger, it connects to WhatsApp, Telegram, Discord, and 50+ other services.

The key difference from ChatGPT or Claude:

  • ChatGPT: "Here's how you could organize your files..."
  • OpenClaw: silently renames, sorts, and moves 47 files into 5 folders "Done."

That's the gap between a chatbot and an agent. OpenClaw bridges it.

Step 1: Installation (~5 minutes)

You need Node.js 18+ and an API key (Claude or OpenAI).

# One-liner install
curl -fsSL https://openclaw.ai/install.sh | bash

# Or via npm
npm install -g openclaw

# Verify
openclaw --version
Enter fullscreen mode Exit fullscreen mode

Then run the setup wizard:

openclaw onboard
Enter fullscreen mode Exit fullscreen mode

This walks you through: choosing a model provider, entering your API key, connecting a chat platform (I picked Telegram), and setting sandbox permissions.

Start with sandbox mode. Full access mode lets OpenClaw run any shell command on your machine. You want to trust it first.

Step 2: Connect a Chat Platform (~10 minutes)

This is what makes OpenClaw different from Claude Code or Cursor. You message it from your phone and it executes on your computer.

Telegram setup (what I use):

  1. Message @botfather on Telegram, send /newbot
  2. Copy the bot token
  3. Configure it: 
openclaw config set channels.telegram.botToken "YOUR_TOKEN"
openclaw config set channels.telegram.enabled true
Enter fullscreen mode Exit fullscreen mode

Now you can text your bot "deploy the staging branch" from the train and it happens. I've been running Claude Code sessions from my phone while walking the dog.

Step 3: Fix the Memory Problem (Critical)

This is where most people give up. You install OpenClaw, have a great conversation, come back the next day — and it has zero memory of what happened.

The default memory is basically a daily log file. It's not enough.

The three-tier memory model that actually works:

Tier What It Stores Where Lifecycle
Information Raw notes, conversation logs memory/learning/ Append-only, search when needed
Knowledge Daily summaries, key decisions memory/YYYY-MM-DD.md One file per day (built-in)
Wisdom Core methods, long-term patterns MEMORY.md Keep under 100 lines, loaded every session

The bottom tier is what matters most. Your MEMORY.md gets loaded at the start of every conversation. Keep it concise — mine has my coding preferences, project context, and a few hard-learned rules.

The seven core files (each one has exactly one job):

  • AGENTS.md — how to work (workflows, procedures)
  • SOUL.md — personality and principles
  • USER.md — your preferences and decision patterns
  • TOOLS.md — tool configs and usage guides
  • MEMORY.md — long-term knowledge (under 100 lines)
  • ERRORS.md — mistakes made and lessons learned
  • SHARED.md — shared context across multiple agents

Iron rule: one piece of information lives in exactly one file. I once had the same rule duplicated in four files. Updated three, missed one. The agent got confused and started contradicting itself.

The Write vs Edit Disaster

My worst early mistake: I had 345KB of learning notes (8,509 lines). A scheduled task used write instead of edit to update it. Write overwrites the entire file. One month of notes replaced by 9.9KB.

The fix that's now in every agent's ERRORS.md:

Never use write on existing files. Always use edit to append.

Step 4: Build a Search Decision Tree

OpenClaw has multiple search tools: web_fetch, curl, browser, plus third-party skills. Without guidance, it picks randomly and wastes tokens on trial-and-error.

I wasted hours watching it try web_fetch → fail (SSRF block) → try to reconfigure → crash → try curl → succeed. Same mistake, every time.

The decision tree I use:

Does it need JS rendering or login?
├── Yes → Use browser
└── No → Use default search skill
    └── Fails?
        ├── SSRF error → curl
        ├── Other error → web_fetch
        └── All fail → browser (last resort)
Enter fullscreen mode Exit fullscreen mode

Special rules:

  • GitHub search → browser only
  • Your own repos → gh CLI
  • Sites that block your IP → browser with proxy

Write this into SHARED.md. Every new agent reads it on startup. One agent's mistake becomes every agent's knowledge.

Step 5: Solve the Context Window Problem

This one is subtle. OpenClaw is doing a complex task, gets halfway through, then asks: "What were we working on again?"

Not a bug. The context window fills up, OpenClaw auto-compresses old messages, and your original instructions get summarized away.

The fix: plan files.

Before any multi-step task, tell OpenClaw to write a PLAN.md:

1. Task description and success criteria
2. Step-by-step breakdown
3. Current progress (updated after each step)
4. Decisions made and why
Enter fullscreen mode Exit fullscreen mode

When context resets, the first thing the agent does is read PLAN.md. It picks up exactly where it left off.

I use this for everything longer than 3 steps. Writing code, researching topics, setting up infrastructure. The plan file is the agent's external brain.

What I Actually Use It For

After two weeks, here's my real workflow:

  • Morning: OpenClaw runs a heartbeat task every 6 hours — checks my servers, summarizes overnight emails, updates my daily log
  • During work: I message it on Telegram for quick tasks — "run tests on the dev branch", "check if the SSL cert is expiring"
  • Coding: Still use Claude Code / Cursor for in-IDE work. OpenClaw orchestrates around them
  • Evening: It consolidates the day's notes into my memory system automatically

Honest Downsides

Not for non-developers. You need terminal comfort, API key management, and patience for configuration.

Can be too autonomous. Early on it decided to "help" by reorganizing my project directory. Start with sandbox mode.

Token costs add up. The heartbeat system and memory consolidation use tokens even when you're not actively chatting. Monitor your API spend for the first week.

Young project. Documentation is patchy. The Discord community is your best resource.

Is It Worth the Setup Time?

If you spend more than 30 minutes a day copy-pasting between AI chat windows and your actual tools — yes. The two hours of setup pays for itself in the first week.

OpenClaw is free, open source, and 68K+ developers are already using it. Install it, spend an afternoon configuring the memory system, and you'll have an AI assistant that actually assists.

curl -fsSL https://openclaw.ai/install.sh | bash
Enter fullscreen mode Exit fullscreen mode

I test AI tools and write about what actually works. More at OpenAIToolsHub — 100+ independent AI tool reviews.

Top comments (0)