Kimi Claw
Moonshot AI’s managed OpenClaw — one-click deployment of the open harness as a 24/7 hosted agent, preconfigured with Kimi K2.6 Thinking and ready-made skills, reachable across messaging apps. Currently previewing Claw Groups. Existing self-hosted OpenClaw instances can be linked rather than replaced.
Three deployment targets, and the third is the unusual one: a cloud server (isolated data, always on), your own desktop (local file access), or an idle Android phone.
A third kind of vendor on top of OpenClaw
openclaw records the pattern: the project supplied gateway-first breadth and left a governance gap that platform vendors filled — NVIDIA with an external sandbox and policy runtime, Microsoft with Windows containers and Entra identity. Both are infrastructure vendors adding containment.
Moonshot is a model lab, and it adds something else entirely: hosting and a model. There is no policy story here, no identity layer, no sandbox — the pitch is that setup disappears and the agent is already wired to a good model. So the layer above OpenClaw now splits by what the vendor already sells: silicon and runtimes get containment, an operating system gets identity, and a model lab gets managed hosting with its weights inside.
That makes it the clearest instance yet of a lab integrating downward — from the model into the
harness that consumes it — and the mirror of ../llm-providers-wiki’s sakana-fugu, which sells
an agent system as a model. One collapses the model/product boundary from each side.
It ships K2.6, not K3
Worth recording because the timing is exact: kimi-k3 released 2026-07-27, and this product runs K2.6 Thinking the following day. The page gives no reason. Plausible candidates — an always-on agent has a very different cost profile from a flagship, and productizing a day-old 2.8T model takes longer than announcing it — but the source supports none of them, so this is recorded as an observation rather than an explanation. It’s a useful reminder that a lab’s flagship and the model its own products actually run can differ.
Tier
T3 — a product landing page. No pricing, no availability detail, no named skills, no description of what Claw Groups actually does beyond the name, and nothing about how the hosted instance is isolated or governed. Everything here is what the vendor chose to put on a marketing page, and the governance silence is conspicuous given that OpenClaw’s known weak spot is ClawHub skill security (an audit found 341 malicious skills) — a hosted deployment with preinstalled skills inherits that question and the page doesn’t address it.
Related
openclaw · nemoclaw · microsoft-scout · agent-skills · agent-guardrails · kimi-k3 · hermes-agent · synthesis