220,651
GitHub stars in 84 days

OpenClaw Deep Dive

How the fastest-growing open-source project in history actually works

A case study built with the PM AI Partner Framework

Ahmed Khaled Mohamed · February 2026 · Use arrow keys to navigate

What Is OpenClaw?

A self-hosted AI agent that runs on your machine and acts through your messaging apps

Unlike ChatGPT or Claude, OpenClaw doesn't just answer questions — it executes actions. Book flights. Manage email. Run terminal commands. Browse the web. Control smart home devices.

It connects through WhatsApp, Telegram, Slack, Discord, Signal, and iMessage — no new app to install. Your data stays local. No subscription, no cloud.

Created by Peter Steinberger (founder of PSPDFKit). Launched November 2025. MIT licensed.

💬

6 messaging platforms

WhatsApp, Telegram, Slack, Discord, Signal, iMessage

🏠

100% local

Runs on your machine, data never leaves

🔧

3,000+ skills

ClawHub marketplace, community-built

🤖

Autonomous

Heartbeat acts while you're away

Architecture

Four components, one control plane

WhatsApp
Telegram
Slack
Discord
Signal
iMessage
⚡ Gateway
Node.js WebSocket hub on 127.0.0.1:18789 — routes messages, manages sessions, enforces security, loads plugins
🧠 Brain
LLM reasoning (Claude, GPT, Grok, Ollama)
🤲 Hands
Shell, Files, Browser, HTTP
💾 Memory
Local Markdown in ~/.openclaw/memory/
💗 Heartbeat
Every 30 min: checks tasks, acts autonomously

Star Growth: 0 → 220K in 84 Days

The fastest-growing software repository in GitHub history

Nov 240 starsRepo created
Jan 24~1,000Organic growth
Jan 26+25,310Single-day record
Jan 27~40,000Anthropic trademark
Jan 30~60,000OpenClaw rebrand
Feb 8~145,0001,000+/day average
Feb 17200,00084-day milestone
Feb 23220,651Current

Ecosystem & Composition

370 contributors, 42K forks, 50 releases, 5 language reimplementations

ProjectLanguageFocus
OpenClawTypeScriptOriginal
ZeroClawRustPerformance
PicoClawGoLightweight
NanoClawPythonML integration
TinyClawShellMinimal

ClawHub Marketplace

3,000+ community skills · 341 confirmed malicious (~11%)

The Wildest Week in Open Source

Three names, one crypto scam, $16M in fake tokens

November 2025
Clawdbot launches — a play on Anthropic's "Claude"
January 27, 2026
Anthropic sends trademark request. Renames to Moltbot — "same lobster soul, new shell"
January 28, 2026
Scammers grab abandoned accounts within 10 seconds. Launch $CLAWD token on Solana.
Hours later
$CLAWD hits $16M market cap. Crashes 90% when Steinberger denies involvement.
January 29, 2026
Final rebrand to OpenClaw. No abandoned accounts this time.

The Lesson

In 2026, a naming decision on an open-source project can trigger a $16M financial event in under 24 hours. The intersection of open source, crypto speculation, and social media impersonation created a new category of risk.

By the Numbers

Name changes3
Time to hijack accounts~10 seconds
Fake token market cap$16M
Token crash-90%

The Security Wake-Up Call

6 CVEs, 900 malicious skills, 40,000 exposed instances

CVESeverityRisk
CVE-2026-25593CriticalUnauthenticated command execution
CVE-2026-25253High 8.8One-click RCE via malicious URLs
CVE-2026-26323HighCI/supply-chain injection
CVE-2026-26327MediumGateway impersonation (mDNS)
CVE-2026-26317MediumPrivilege escalation
CVE-2026-26329MediumToken replay attacks

The "Lethal Trifecta"

1. Access to private data (files, emails, messages)
2. Exposure to untrusted content (web browsing)
3. External communication (send messages, API calls)

A compromised agent can exfiltrate data through channels you'd never check.

~900
Malicious skills (20% of ClawHub)
40,000+
Exposed instances (35% vulnerable)

Why OpenClaw Exploded: 3 Product Insights

The innovation isn't the AI model — it's the interaction model

01

Local-First Is the Killer Feature

In an era of subscription fatigue, "runs on your machine, data never leaves" is a powerful proposition. Users bought so many Mac Minis as dedicated agent machines that Apple stores sold out.

Removing cloud dependency is a feature, not a limitation.

02

Messaging-as-Interface

OpenClaw meets you in WhatsApp, Telegram, Slack — apps already open on your phone. No new UI to learn. Distribution through existing channels beats purpose-built interfaces.

The best interface is the one you already use.

03

The Heartbeat Creates Attachment

Every 30 minutes, the agent checks for tasks and acts autonomously. This transforms OpenClaw from a tool you use into an assistant that works for you. Chatbots respond. Agents initiate.

The smallest feature that creates the largest behavioral shift.

How This Was Built: The Framework in Action

4 skills, ~3 hours, from zero knowledge to complete analysis

/technical-analyst

Map the Architecture

"Trace a WhatsApp message through to an executed action."

Output: Architecture diagram, component breakdown

/data-analyst

Collect the Numbers

"Get star history, contributor stats, language breakdown."

Output: Chart-ready datasets, growth milestones

/devil-advocate

Challenge the Narrative

"What are the real risks? Where does this fall apart at scale?"

Output: CVE analysis, permission model critique

/builder

Build the Deliverables

"Create an interactive presentation with charts, plus a PPTX deck."

Output: This presentation, PowerPoint deck, article

The Compounding Effect

Technical analysis told us what to measure. Data collection revealed the growth inflection points.

Devil's advocate found the security counter-narrative. Builder mode turned all of it into shareable formats.

Structured thinking compounds. Each mode contributes something the others can't.

Key Takeaways

01

You don't need to be a developer to understand complex technical projects. You need a structured approach — architecture, data, risks, communication.

02

OpenClaw's innovation is the distribution, not the AI. Messaging-as-interface, local-first, autonomous heartbeat. The model is swappable; the interaction model is the moat.

03

Speed and security are inversely correlated at scale. 220K stars = 900 malicious plugins + 6 CVEs + 40K exposed instances.

04

The open-source AI agent era has new risk categories. When an agent has your full system permissions, the blast radius is your entire digital life.

05

AI works best as a thinking partner when you give it structure. Agent modes aren't magic — they're forcing functions that ensure multiple angles.

Framework: github.com/ahmedkhaledmohamed/PM-AI-Partner-Framework