OpenClaw vs. Hermes: Unified AI Agent Definition, Divergent Control Mechanisms

At Microsoft’s Build keynote, CEO Satya Nadella announced a shift toward autonomous AI agents that run without user initiation. Two open‑source projects, OpenClaw and Hermes Agent, embody this vision but take opposite architectural approaches.

OpenClaw’s gateway‑first design

OpenClaw is an independent open‑source framework launched by Peter Steinberger in early 2025 and renamed in January 2026. Its core is a central gateway that connects an agent to dozens of messaging channels (WhatsApp, Discord, Slack, etc.). The framework includes a runtime environment, a gateway for message I/O, and cross‑session persistent memory, plus tools, identities, extensible skills, and policy/observability controls.

OpenClaw’s public skill marketplace, ClawHub, hosts thousands of community‑contributed skills. By June, the GitHub repository approached 380 000 stars, reflecting visibility rather than production use. OpenAI became a sponsor in February, and Nvidia packaged OpenClaw in NemoClaw at the March GTC conference, adding sandboxing and policy enforcement. Microsoft later integrated OpenClaw natively into Windows execution containers and released Scout, an agent running on the OpenClaw gateway with its own Entra identity and connections to Teams, Outlook, and SharePoint, thereby adding governance and identity management.

“The framework maintains runtime, memory, and governance so agents can operate unattended.”

Hermes’s memory‑first design

Hermes Agent, released by Nous Research under the MIT license on 25 Feb, is written in Python and intended to run persistently on self‑hosted infrastructure (VPS, home servers, laptops). Its core is a hierarchical, cross‑session persistent memory that enables agents to learn developers’ work, develop new skills after difficult tasks, and continuously refine them. Skills follow the agentskills.io standard, making them portable across agents.

Hermes quickly gained traction: GitHub stars surpassed 100 000 in mid‑May and reached about 160 000 by month‑end; on 10 May it topped OpenRouter’s daily token leaderboard with 224 billion tokens, outpacing OpenClaw’s 186 billion. The command hermes claw migrate allows a one‑step import of OpenClaw users’ settings, memory, skills, and keys.

“Developers can retain an agent that carries weeks of code, conventions, and decisions without rebuilding context each morning.”

Head‑to‑head comparison

Both projects agree on the definition of an AI agent but diverge on the primary control point. OpenClaw emphasizes a gateway that unifies multi‑channel communication, while Hermes emphasizes persistent memory that accumulates developer context. This mirrors the classic trade‑off between hosted cloud services (convenient, vendor‑managed) and self‑hosted infrastructure (full control, operational responsibility).

Typical use‑case considerations:

Regulated enterprises needing audit and policy control: OpenClaw (or Microsoft Scout/NemoClaw) offers integrated governance and identity, though it ties customers to the vendor’s stack.

Developers wanting a learn‑and‑portability agent: Hermes provides durable memory and self‑improvement at the cost of managing the underlying infrastructure.

Teams communicating across multiple chat platforms: OpenClaw’s extensive skill market and channel coverage give it an advantage, albeit with variable skill quality and supply‑chain risk.

In practice, deployments often combine both, selecting the component that best fits the workload.

Why the control layer matters

Security teams can now restrict which folders an agent may read, avoiding the broad permissions of early OpenClaw deployments. Audits of OpenClaw’s skill market uncovered 341 malicious entries, highlighting the need for governance.

“Nvidia and Microsoft are racing to provide governance, identity, and observability for any agent a customer chooses.”

Both vendors aim to embed these controls in the runtime layer, making it more durable than any single infrastructure model.

Future outlook

The AI‑agent market is moving from model selection to runtime, governance, and memory layers. OpenClaw demonstrates that a broad entry point and large skill ecosystem attract developers and major players like OpenAI, Nvidia, and Microsoft. Hermes shows that persistent memory and self‑improving skills can sustain heavy daily usage without vendor‑specific platform support. Whether breadth and depth will remain in separate projects is uncertain, but ownership of memory, runtime, and governance will become the decisive competitive factor.

Key takeaway

Enterprises must understand who controls an agent’s accumulated memory, which tools it can invoke, and who owns the runtime environment, as these factors dictate lock‑in costs and security responsibilities.