← Back to content hub

How OpenClaw was created:
From an hour of code
to 247,000 GitHub stars

One hour. A WhatsApp connection. A Claude API. And the question: Why doesn't this exist yet? This was the beginning of OpenClaw - the open source AI agent that shook the developer community in 2026.

Author: Joshua Heller
Published: April 2026
Reading time: approx. 16 minutes
Category: Product & Technology
1h

01 —The original idea: “WhatsApp Relay”

Peter Steinberger sat at his computer one evening at the end of November 2025 and a simple question arose: “Can I give my computer a task via WhatsApp - and he does it while I'm somewhere else?"

The answer was: Technically yes, but no ready-made tool existed for it. So he built it himself. He called it internal “WhatsApp Relay”. The mechanics were radically simple: a message goes to a WhatsApp number. The WhatsApp API forwards them. Anthropic's Claude API receives them, thinks about it, returns an instruction. The instruction will executed on the local computer.

First working version: one hour.

“He had a deceptively simple question: Can an AI assistant make progress check his work via a chat app? One evening he bandaged one Messaging API using Claude's API, gave the system local system access - and had a prototype in an hour."
— emergent.sh, Complete Guide OpenClaw 2026

Steinberger thought it was so obvious that OpenAI or Anthropic would soon build exactly that. So he put the code on GitHub – not as a product, but as an experiment. The project was called Clawdbot, a pun on "Claude" with a lobster claw as a mascot. The internet responded with an enthusiasm he didn't expect.

02 —What OpenClaw really is - and what it isn't

Above all else, one thing must be clear: OpenClaw is not a chatbot. It is also not a model, not a SaaS solution and not a competing product to Claude or ChatGPT.

OpenClaw is one Orchestration layer – a framework, that runs on your own computer and from any AI model (Claude, GPT-5, Gemini, DeepSeek, Llama via Ollama) one capable of action agents makes.

💡 The simplest explanation

ChatGPT answers questions. OpenClaw gets things done. It reads emails, books appointments, executes terminal commands, scrapes websites, fills out forms, manages files – and does this autonomously, on your own Hardware, 24/7. It is aptly called: "Claude with hands."

The crucial difference lies in the ratio of input to output. With a chatbot, you enter a prompt and receive text back. With OpenClaw you specify a goal - and the agent decides independently She carries out which steps are necessary and observes the result and iterates until the task is completed.

The project is MIT licensed, free and open source. Written in TypeScript and Swift. Runs locally on your own hardware – Data only leaves your own computer if you use an external one AI model (Claude API, GPT API) is used.

03 —Architecture: The Four Layers of the Agent

OpenClaw does not work like a conventional software system solid logic. It is a Composition system, that the agent made up of individual building blocks every time you start – from the file system, in real time.

The architecture consists of four central layers:

🏗 OpenClaw Core Architecture
identity
SOUL.md + IDENTITY.md
Personality, values, communication style, specific role – injected into the system prompt at every start
memory
MEMORY.md + Vector Index
Persistent long-term memory in Markdown files, semantically searchable via SQLite + embeddings
skills
Skills + HEARTBEAT.md
Modular Capabilities (Markdown + Script), Proactive Scheduled Actions, ClawHub Marketplace
Gateway
Harness / MCPorter / Messaging Bridge
Connection to messaging apps (WhatsApp, Telegram etc.), MCP protocol, running on the local system

The crucial difference to other agent frameworks: Standard agents are amnesiacs. Start on every reboot them at zero. OpenClaw solves this by consistently reading configuration files at startup - the agent "wakes up" and literally reads itself into its existence into it.

A deep understanding of your own system is also built in: The agent knows its own source code, its configuration, which models he uses and which options are available. This creates what Steinberger “System Awareness” called – not in a philosophical sense, but in a highly practical one: An agent that knows its own architecture can know itself develop and create new skills on demand.

04 —SOUL.md – The “character” of the agent

One of OpenClaw's most innovative concepts is the SOUL.md file. It is the foundation of every OpenClaw agent - a simple Markdown file, which describes in natural language who the agent is, how he communicates, what his values are and what boundaries he has to adhere to.

📄 What SOUL.md contains

Typical SOUL.md sections: Identity (Who are you?), Communication style (How do you communicate?), Core Values (What do you stand for?), Non-negotiables (What do you never do?). The file is injected into the system prompt every time the agent is started. No database access, no API – just Markdown.

The philosophical difference to an ordinary system prompt lies in the intention: "You're not a chatbot. You're becoming someone." This is how a much-quoted SOUL.md example from the community begins. The first line of a standard prompt would be: "You are a helpful assistant." This is a behavior instruction. SOUL.md is an identity definition.

Along with IDENTITY.md (specific role), MEMORY.md (knowledge) and HEARTBEAT.md (planned autonomous actions) forms SOUL.md that, what Steinberger did? "Harness" called – the framework that ensures the agent is always running, is always available and always across all channels carries the same context.

05 —The memory system: Persistence without a database

The most common criticism of AI chatbots is: "Goldfish Brain." As soon as the browser window is closed, everything is forgotten. OpenClaw solves the problem in a radically simple way - and that's exactly where the elegance lies.

Memory in OpenClaw are Markdown files on the local hard drive. The agent is instructed to provide relevant facts after each meaningful interaction to write into his memory files. The next time it starts, it reads it. “Mental notes” disappear when restarting – files do not disappear.

⚙️ Technical: How retrieval works

OpenClaw is building one Vector index about the memory files: Chunking (approx. 400 tokens, 80 token overlap) → Embedding via OpenAI, Gemini or local GGUF model → SQLite with optional sqlite-vec acceleration. When retrieved, it combines Vector similarity (semantic match) with BM25 keyword relevance (exact token match). Score: vectorWeight × vectorScore + textWeight × textScore. The result: An agent that understands "Mac Studio gateway host" even if the note only says “the machine on which the gateway is running”.

06 —Skills: How OpenClaw expands its capabilities

In most agent frameworks, a "tool" is a JSON function signature. A rigid interface. OpenClaw thinks radically differently: Skills are Markdown files with embedded scripts.

To teach OpenClaw a new skill, you create a directory writes a Markdown file that describes what the skill does, adds the script – and the agent can use it immediately. No reboot, no code deployment, no API registration.

📧
Email & Calendar
Read, categorize, reply and forward incoming emails. Book appointments, identify conflicts, set reminders.
🖥
System control
Execute terminal commands, read/write/move files, start processes, analyze logs - with root level access.
🌐
Web automation
Control browsers, fill out forms, scrape websites, extract data - without manual intervention.
🤖
Sub-agents (Lobster)
Specialized sub-agents spawn and coordinate. A master agent delegates subtasks – parallel execution.
Proactive actions
HEARTBEAT.md defines planned actions: daily reports, monitoring checks, automatic follow-ups - without user triggers.
🛠
Self-extension
If a feature is missing, the agent writes the required script itself. The WhatsApp voice memo example is canonical.

The ClawHub Marketplace is the app store for OpenClaw skills. Thousands of community-developed skills cover everything: from Notion integration to bright data scraping to Spotify control. The catch: Quality varies greatly, and before security hardening In 2026, there were 341 malicious skills on ClawHub with a contamination rate found by 12% - a direct indication of why enterprise deployments curated, verified skill sources.

The supported messaging platforms (as of April 2026):

Whatsapp Telegram Discord signal Slack iMessage matrix LINE QQBot Microsoft Teams

All channels share a single, shared memory. The agent is aware of a command via WhatsApp in the morning, when asked via Slack in the evening.

07 —Product evolution: From v0.1 to NemoClaw

Nov. 2025 · Phase 0
“WhatsApp Relay” – The first hour
Peter Steinberger connects WhatsApp with the Claude API. Local system execution. First version in an hour. The project is internally called “WhatsApp Relay” – no name yet, no GitHub repo yet.
WhatsApp ↔ Claude API1 hour buildPrivate
Nov. 2025 · Phase 1
Clawdbot – First public release
Published on GitHub under the name Clawdbot. Mascot: Cartoon lobster “Clawd”. Support for multiple AI models. First plugin architecture. Browser automation is added – the jump from a simple chatbot to an automation agent.
Multi-modelPlugin architectureBrowser automationOpen source
Jan 25, 2026 · Phase 2
Official launch – 9,000 stars in 24 hours
Official GitHub launch with global media coverage. 9,000 stars in 24 hours. Andrej Karpathy calls it the most impressive thing he's ever seen. 2 million visitors in one week.
9K Stars / 24h2M weekly visitors
27–30 Jan 2026 · Phase 3
Moltbot → OpenClaw: Two renames in one week
Trademark Complaint by Anthropic. 5am ​​Discord Brainstorm. Moltbot. Handle sniping. $CLAWD crypto crash. Three days later: OpenClaw. Trademark checked, domains secured. (Full naming history →)
Trademark disputeOpenClawMIT license confirmed
Feb. 2026 · Phase 4
100,000 Stars · Security Hardening · OpenAI Hire
100K Stars – fastest growth in GitHub history. Security researchers report 1,000+ unsecured instances. Steinberger publishes 34 security commits. ClawHub Marketplace is growing. Vercel sponsors the project. February 14: Steinberger moves to OpenAI, OpenClaw to independent foundation.
100K starsClawHubFoundation
March 2026 · Phase 5
247,000 stars · NVIDIA NemoClaw · Global ecosystem
NVIDIA announces NemoClaw at GTC 2026 – an enterprise version with GPU acceleration. Tencent, Alibaba and ByteDance are building platforms on OpenClaw. 300,000 lines of code. 47,700 forks. Feishu/Lark support for the Chinese market.
247K starsNVIDIA NemoClaw300K+ LoC

08 —OpenClaw vs. ChatGPT Enterprise vs. n8n – the honest comparison

criterion OpenClaw ChatGPT Enterprise n8n/Zapier
type Autonomous agent + framework Chatbot with restricted Actions Workflow automation
Acts autonomously ✓ Yes – multi-stage, without confirmation ⚠ Restricted – Tools via API ✗ No – rule-based, no AI thinking
Data sovereignty Local, on-premise possible OpenAI US Cloud Depends on hosting
memory Persistent, across sessions Memory feature, limited No AI memory
Skill expansion Markdown files, self-writing GPT plugins (limited) Lots of connectors
Multi-model Claude, GPT, Gemini, DeepSeek, Ollama OpenAI only About AI nodes
Messaging native WhatsApp, Telegram, Slack and much more. Browser/API only About webhook connectors
Open source MIT license Closed Source n8n: partially, Zapier: no
Enterprise ready (out of box) Needs professional setup Can be used immediately For simple flows
Cost € – LLM API costs only €€€ – Subscription + API € to €€ – subscription model

The short version for decision-makers: n8n and Zapier are the right choice for rule-based, predictable processes without AI thinking. ChatGPT Enterprise is the easiest way to use AI, with the disadvantage of US cloud dependency and no real agent logic. OpenClaw is the most powerful tool for complex, Autonomous workflows – but professional deployment is needed in order to Meet enterprise requirements.

09 —What companies in the DACH region need to know

OpenClaw is not a plug-and-play enterprise product. The security researchers at SOCprime found shortly after the launch over 1,000 unsecured instances on the public internet – with exposed API keys, chat histories and system credentials. The reason: The agent needs root-level access to the system, to fulfill his tasks. Without a professional security concept This is a massive attack surface.

⚠️ Enterprise requirements for OpenClaw

What every business installation needs: Sandboxed Execution Environment Role-Based Access Control (RBAC) · Audit logging of all agent actions · Encryption of all communication channels · Curated, verified skill sources instead of ClawHub community · AVV contract for all API calls to external models · On-premise deployment for GDPR-critical data

The good news: Enterprise maturity is growing rapidly. NVIDIA NemoClaw addresses GPU acceleration and enterprise governance. The OpenClaw Foundation works on team management, compliance tools and auditable logging. And focus on projects like IronClaw (NEAR AI). explicitly focuses on safer sandboxing.

For companies in the DACH region that want to start using OpenClaw today, There is no way around an experienced implementation partner – someone who knows enterprise architecture, GDPR-compliant deployments has carried out and understands where the limitations of the current framework lie.

— — —

FAQ: Frequently asked questions about OpenClaw

How was OpenClaw built?

Peter Steinberger built the first version in an hour: He connected the WhatsApp API with Anthropic's Claude API and made it possible to delegate tasks to the AI ​​model on the local computer. The project was published on GitHub in November 2025 and has since grown by the community to 300,000+ lines of code.

What is the difference between OpenClaw and ChatGPT?

ChatGPT is a chatbot that generates text. OpenClaw is an execution framework that performs autonomous actions in the real world: reading emails, executing terminal commands, managing calendars, controlling browsers, moving files - without confirmation at each step. It runs locally, is model-agnostic and open source.

What is SOUL.md?

SOUL.md is a Markdown file that defines the personality, values, and behavior rules of an OpenClaw agent in natural language. It is injected into the system prompt every time the agent is started and ensures a persistent, consistent identity across all sessions and messaging channels.

Which AI models does OpenClaw support?

OpenClaw is model-agnostic: It supports Claude (Anthropic), GPT-5 (OpenAI), Gemini (Google), DeepSeek, KIMI K2.5, Xiaomi MiMo as well as locally operated models via Ollama. Companies can choose any model – or combine different models for different tasks.

Is OpenClaw suitable for businesses?

OpenClaw has enormous potential for companies, but requires a professional enterprise architecture: sandbox environment, role-based access control, audit logging and GDPR-compliant deployment. A standard community setup is not enterprise-ready. Clawgency implements OpenClaw with this architecture for companies in the DACH region.

What is the ClawHub Marketplace?

ClawHub is the official skill marketplace for OpenClaw - comparable to an app store for agent skills. Thousands of community-developed skills expand the agent. For enterprise use, only verified, curated skills should be used; Malware-infected skills were found in the community pool.

🦀 OpenClaw experts in the DACH region

Make OpenClaw enterprise-ready – for your company

Clawgency implements OpenClaw with professional enterprise architecture: GDPR compliant, secure, scalable – live in 2-4 weeks, with 90 day ROI guarantee.

Book a free initial consultation →