14.1 OpenClaw Panorama: From Clawdbot to OpenClaw

🦞 "How a weekend project became a phenomenon — and why its name changed three times."

Why OpenClaw Matters

Before 2025, Agents were mostly demos (AutoGPT, BabyAGI) stuck in a CLI or a crude web UI. The thing that pushed Agents into daily life was the action-oriented Agent living in chat apps — you @ it in WhatsApp and it searches, edits files, runs commands.

OpenClaw is a leading representative. Its GitHub growth was unprecedented — hundreds of thousands of stars within months. How did it start? What fundamentally separates it from an industrial harness like Claude Code? This section gives the panorama.

A Brief Naming History

The same codebase changed names three times in under two months:

What this history teaches:

1. Naming is engineering — a wrong name brings recurring trademark/SEO/community costs.
2. Rewrites flourished — nanobot (Python ~4000 lines), ZeroClaw (Rust), NanoClaw (Go + Apple containers), IronClaw (Rust + WASM sandbox), NullClaw (Zig, 678KB static binary) all use OpenClaw's core architecture as a starting line — itself a "collective quality review."
3. OpenClaw chose the community-driven route — the biggest divide from Claude Code (closed source).

OpenClaw's Place in the Harness Spectrum

Action-oriented Agent Harness spectrum (2026)
│
├── Multi-channel · consumer · community-driven (open source)
│   └─ OpenClaw (Ch.14) — put the Agent in chat apps
│
├── Multi-channel · consumer · self-evolving (open source)
│   └─ Hermes Agent (Ch.15) — Agent writes its own skills
│
├── Terminal IDE · industrial · closed (source-available)
│   └─ Claude Code (Ch.16) — the coding IDE
│
└── Terminal IDE · developer workshop · everything-is-a-plugin (open source)
    └─ DeepSeek Harness (Ch.17) — model-agnostic + swappable kernel

The "user key" difference is decisive:

📌 OpenClaw moved the user key from CLI to chat app — a structural change: inputs become shorter and more colloquial, demanding stronger context summarization and ambiguity tolerance.

Three Fundamental Differences from Claude Code

Key takeaways: scene (not capability) is the dividing line; license ≠ safety; entry point determines design — a chat entry means "messages must fit in ~1.6K chars," driving summarization and task-decomposition submodules that a CLI entry doesn't need.

The 4 Core Subsystems

Gateway (message hub)

Normalizes all channels into IncomingMessage/OutgoingMessage. The flow across the four subsystems is:

Agent Loop — streaming inference → tool-call parsing → execution → write-back; with compression and permission checks.

Toolbox — file / shell / web / message / calendar / skill / memory tools for daily tasks.

Skills & Plugins — SKILL.md (Markdown + optional script) registers new capabilities; the format is compatible with Claude Code.

💡 These four map directly onto Chapter 8's "six engineering pillars": Gateway ⊂ Channels; Agent Loop ⊂ loop + context; Toolbox ⊂ tools; Skills ⊂ skill system.

Section Summary

Next section: 14.2 Installation & 4 Deployment Modes