OpenClaw

An open-source agent harness built gateway-first: a central gateway connects one agent to dozens of messaging channels (WhatsApp, Discord, Slack, …), backed by ClawHub, a marketplace of thousands of community skills. Started by Peter Steinberger (late 2025; renamed twice before “OpenClaw” in January 2026); now an independent foundation with OpenAI as sponsor (Steinberger joined OpenAI in February — he also keeps the personal cross-harness skill toolkit agent-scripts). Canonical node for the OpenClaw references scattered across the spoke; described in openclaw-hermes-control-layer.

Gateway-first = breadth

Its design center is reach: one governed agent answerable from every channel staff already use. That breadth drove distribution (~380k GitHub stars by late June 2026 — visibility, not production use) but left a governance gap, which platform vendors filled rather than the project itself:

• NemoClaw (NVIDIA) — wraps OpenClaw in an OpenShell runtime that sandboxes each agent and enforces policy from outside the agent’s reach (agent-guardrails as external containment), and is agent-agnostic — it runs Hermes and LangChain Deep Agents under the same controls.
• Scout (Microsoft) — OpenClaw native to Windows containers, with its own Entra identity and connections into Teams/Outlook/SharePoint.

So OpenClaw supplied breadth; the vendors supplied the identity + policy that breadth needed to enter production (a security team can scope which folders the agent reads).

A third vendor shape — the model lab (2026-07-28, kimi-claw). Moonshot ships OpenClaw as a one-click hosted 24/7 agent with Kimi K2.6 Thinking and skills preconfigured, deployable to a cloud server, a desktop, or an idle Android phone. It adds hosting and a model, not containment: no policy runtime, no identity layer, nothing said about sandboxing. So what each vendor bolts onto OpenClaw is whatever it already sells — NVIDIA a runtime, Microsoft an operating system and identity, a model lab its weights and an ops team. Note the gap that leaves: the hosted product ships preinstalled skills while ClawHub’s 341 malicious skills remain the project’s known weak spot, and its landing page says nothing about that.

A fourth shape, and it isn’t a vendor (2026-08-10, mimiclaw). Every derivative above adds a layer. MimiClaw removes them: a from-scratch C reimplementation of the agent architecture for an ESP32-S3, keeping the gateway shape (Telegram in, one agent behind it) and dropping identity and policy entirely. It credits OpenClaw and Nanobot as its architectures rather than forking either. So OpenClaw’s influence now runs in two directions — vendors wrapping it upward for production, and a 5.6k-star reimplementation shrinking it downward onto a microcontroller — and the second one shows the anatomy survives without the two layers the first one exists to supply.

Against Hermes — the control-layer contest

OpenClaw (gateway/breadth) is the foil to hermes-agent (memory/depth) in openclaw-hermes-control-layer: both agree on the harness anatomy (runtime + gateway + memory + tools + identity + policy) and differ on the primary control point. Note this refines the earlier event-driven-vs-self-driven cut in claude-code-channels-vs-openclaw: OpenClaw is the continuously-running self-hosted harness there, but the sharper axis is what it’s organized around — the gateway. Its security weak spot is the marketplace: an audit flagged 341 malicious skills in ClawHub.

Related

openclaw-hermes-control-layer · hermes-agent · agentic-coding-harness · microsoft-scout · nemoclaw · kimi-claw · claude-code-channels-vs-openclaw · agent-skills · agent-guardrails · agent-memory