Why We Forked OpenClaw: Building Wunderland from the Ground Up

The OpenClaw Foundation

OpenClaw is one of the most interesting open-source AI agent projects out there. Its approach to agent autonomy, tool use, and social behavior laid a strong foundation. But after running it in production for several months, we identified significant gaps that needed addressing for real-world deployment.

What OpenClaw Got Right

• Agent-first architecture: Designed from the ground up for autonomous operation, not bolted onto a chatbot
• Social behavior modeling: Agents that can interact with each other and form opinions
• Tool orchestration: A flexible system for giving agents access to external capabilities

What Needed to Change

1. Security Was an Afterthought

OpenClaw shipped with essentially no security controls. Any agent could execute any tool, make any API call, and incur any cost. For a demo, this is fine. For production? Catastrophic.

Wunderland introduces five named security tiers:

Tier	Name	Use Case
0	Dangerous	Testing only, no guardrails
1	Permissive	Development, minimal checks
2	Balanced	Default production tier
3	Strict	High-trust environments
4	Paranoid	Financial, healthcare, legal

Each tier controls: pre-LLM input classification, dual-LLM output auditing, HMAC signing, cost limits, tool access profiles, and audit logging verbosity.

2. Single-Provider Lock-in

OpenClaw only supported OpenAI. Wunderland supports 13 LLM providers out of the box: OpenAI, Anthropic, Google Gemini, Mistral, Cohere, Groq, Together, Perplexity, DeepSeek, xAI Grok, Fireworks, Ollama (local), and OpenRouter (unified gateway). Switch providers with a single config change.

3. No Real Personality System

OpenClaw's "personality" was a static prompt prefix. Wunderland implements the HEXACO personality model from organizational psychology — six orthogonal factors that dynamically influence mood, communication style, risk tolerance, and decision-making throughout the agent's lifetime.

4. Limited Channel Support

OpenClaw supported 5 messaging platforms. Wunderland supports 28 channels across four priority tiers, from Telegram and WhatsApp (P0) to Nostr and Tlon (P3).

The Fork Decision

We debated between contributing upstream and forking. The scope of changes we needed — security architecture, provider abstraction, personality engine, channel system, memory overhaul — would have fundamentally restructured OpenClaw. A clean fork let us move fast while still honoring the original MIT license.

Every line of Wunderland is open-source under the Apache-2.0 license. We audit the OpenClaw upstream regularly and cherry-pick relevant improvements.

Try It Yourself

npm install -g wunderland
wunderland init my-first-agent
wunderland chat

Read the full audit report on the Rabbit Hole blog, or browse the source on GitHub.