OpenClaw and Imaginclaw solve the same problem — why does it matter where the model runs?
Both are personal AI agent gateways that connect your chat apps (WhatsApp, Telegram, Slack, Discord...) to an AI that can actually act, not just answer. That overlap is what makes comparing them worth doing — this isn't a forced tie-in.
What they have in common
Both go beyond a chat box. Both are self-hosted gateways — you run the process, not a SaaS vendor. Both connect the same kinds of channels to an AI agent that can read context, use tools, and act on a loop, not just reply once.
OpenClaw
A viral open-source personal AI agent — 100,000+ GitHub stars in its first week, built by PSPDFKit's founder. It's model-agnostic: Claude, GPT-4o, Gemini, or a locally-hosted model via Ollama all work interchangeably, using your own API key. It has shell access, browser control, and can send emails on your behalf, in a loop, without asking each time.
Imaginclaw
Paulo's own sovereign assistant, built on Hera — his local LLM runtime. Runs on his own hardware (two GPUs), with local models as the real default path, not an optional mode. Cloud is a deliberate, explicit fallback, not the baseline.
The honest axis
Who controls the model, and what are you actually solving? OpenClaw is open-source in the licensing sense — but by default, your messages and context still transit a third party's cloud (Anthropic, OpenAI, Google), even though the gateway itself runs on your machine. "Open-source" and "sovereign" are not the same claim, and OpenClaw's own docs make that distinction explicit — it's model-agnostic BY DESIGN, not sovereign by default.
"Self-hosted" describes where the gateway process lives. "Sovereign" describes where your data goes. OpenClaw solves the first by design; Imaginclaw was built to solve both.
The real reason this comparison is worth reading
Imaginclaw wasn't built to compete with OpenClaw feature-for-feature. It was built because a self-hosted gateway that still routes every message through a third-party cloud API doesn't actually solve the sovereignty problem — it just moves where the gateway process runs, not where your data goes. If you already run local models for other things, wiring the same assistant pattern on top of them was the more consistent choice, not a novel one.