State of OpenClaw 2026: Enterprise Targets, Anthropic Friction, and 374K Stars

OpenClaw crossed 374,000 GitHub stars and 12 million downloads in May 2026 — a inflection point where it stopped being a scrappy open-source project and became an enterprise target. Here’s the current state of play.

The Anthropic/Claude Code Friction The biggest community story of the past month: Claude Code has been observed scanning repositories for HERMES.md — the OpenClaw agent configuration manifest — and either refusing requests or routing them to a higher-cost “extra usage” billing tier. Users reported cost increases up to 50x. The Hacker News post accumulated 1,336 upvotes and 718 comments. OpenClaw maintainers have not issued a formal response, but the incident has sparked broader discussion about hosted-AI vendor subscription boundaries and agent harness traffic classification.

Enterprise Hardening: NemoClaw NVIDIA shipped NemoClaw alpha — a hardened OpenClaw fork integrated with NeMo guardrails and OpenShell sandboxes. For risk-averse enterprise IT teams, it gives them something defensible in architecture reviews. The fork addresses the same class of TOCTOU and sandbox escape issues disclosed by Cyera in April.

Three Trends Colliding

  • Hosted AI vendors are reclassifying or surcharging agent harness traffic as capacity pressure mounts
  • Government restrictions on autonomous AI agents have been issued in Belgium, China, and South Korea, tightening procurement scrutiny
  • Enterprise-grade hardening of open agents is finally available, shifting the conversation from “let’s pilot a hosted agent” to “what’s our agent control plane?”

The platform is now at the stage where enterprise IT teams need an opinion on OpenClaw this quarter, not next year.

← Back to News