Skip to content
Diadems Tech

Comparison · Hermes Agent and OpenClaw

Hermes Agent vs OpenClaw: two open-source agents, two philosophies

Two open-source AI agents under the MIT licence, two philosophies: the personal assistant on your own hardware, and the agent that writes its own skills. What really separates them, with sources.

Create my agent

OpenClaw, created by Peter Steinberger in November 2025 under the name Clawdbot, is a personal assistant written in Node.js that lives on your machine and talks in your messaging apps. Hermes Agent, released by Nous Research in February 2026, is written in Python and stands out for its learning loop: it writes and corrects its own skills. Both are under the MIT licence, designed for a single trusted person, and need a machine that is always on.

Hermes Agent and OpenClaw, in two paragraphs

OpenClaw was born in November 2025 under the name Clawdbot, a nod to Claude. At Anthropic's request it became Moltbot, then OpenClaw at the end of January 2026. Its creator, Peter Steinberger, joined OpenAI in February; the project is now run by the OpenClaw Foundation, an American non-profit that sells neither a paid offer nor hosting. Its promise: “Your AI assistant, on your own hardware, in every chat app you already use.”

Hermes Agent comes from Nous Research, an American open-source AI lab. Publicly launched in February 2026, it presents itself as “The agent that grows with you”: when it succeeds at a task worth repeating, it writes the procedure into a skill, then improves it with use. Nous Research also sells official hosting, Hermes Cloud.

Hermes Agent and OpenClaw, side by side

Noted on 10 November 2026 from the GitHub repositories and official documentation of both projects. Versions change every week.

OpenClawHermes Agent
PublisherPeter Steinberger, then the OpenClaw FoundationNous Research
LaunchNovember 2025 (as Clawdbot)February 2026
LicenceMITMIT
LanguageNode.jsPython
Release paceSeveral releases a week (2026.11.3 on 10 November)Several patch releases a week (v0.23.1 on 9 November)
Popularity on GitHubAbout 394,000 starsAbout 262,000 stars
MemoryMarkdown files (MEMORY.md, USER.md, daily notes), semantic and keyword searchTwo short files reread in every session, search across past conversations, memory plug-ins (Hindsight, Honcho, Mem0…)
SkillsSKILL.md files, the public ClawHub registrySKILL.md files that the agent writes and corrects itself; can draw on ClawHub and other sources
Messaging appsTelegram, WhatsApp, Slack, Discord, Signal, iMessage, Teams, Matrix, Google Chat and othersThe same or nearly, plus email, SMS, Mattermost, LINE, WeCom and some twenty others
Running commandsOn the host machine, unless a sandbox is configuredOn the machine, in Docker, over SSH or with a sandbox provider
Official hostingNoneHermes Cloud, subscriptions from $20 to $200 a month, inference extra
What it puts forwardYour data and credentials stay on your hardwareThe learning loop: it improves its own skills

Sources: github.com/openclaw/openclaw, docs.openclaw.ai, github.com/NousResearch/hermes-agent, the Hermes Agent documentation, portal.nousresearch.com.

On paper, the two projects do almost the same thing. The difference lies in their intent: OpenClaw is a personal assistant meant to live at your place; Hermes is an agent that builds up know-how and can run on a server, in a container or in the cloud.

Switching from one to the other

Hermes provides a command, hermes import-agent, that takes over the settings, memory and skills of an OpenClaw agent. Secrets only come across on explicit request.

Security: what 2026 has shown

An agent runs commands with your rights. Its security therefore depends less on its code than on where it runs and what it is allowed to install.

OpenClaw

  • A serious flaw at the end of January 2026 (CVE-2026-25253, 8.8 out of 10, rated “high”): a booby-trapped web page could take control of an installation in one click. It was fixed in version 2026.1.29.
  • Malicious skills on ClawHub. In February 2026, Koi Security found 341 out of 2,857, most of them from a single campaign that installed a macOS password stealer; its next count found 824 out of more than 10,700. ClawHub now scans published skills.
  • Thousands of installations left reachable from the internet by their owners: more than 21,000 counted by Censys at the end of January, more than 40,000 by SecurityScorecard in February. It is a configuration problem as much as a software one.

OpenClaw's own documentation says so: treat third-party skills as untrusted code, and be aware that tools run on the host machine for as long as no sandbox is configured.

Hermes Agent

No incident on that scale has been reported. Third-party vulnerability databases do, however, list several vulnerabilities fixed over the course of 2026, and the project is clear about its limits: its security policy states that “The only security boundary against an adversarial LLM is the operating system.” Its command approvals and skill checks are safeguards, not isolation.

And for a team?

Neither is designed to be shared by several people. OpenClaw's security policy states that it does not treat a gateway as a boundary between users who do not trust each other; Hermes's describes it as “a single-tenant personal agent”, in which everyone you authorise is equally trusted.

For a team to share an agent, you therefore have to add a layer: accounts, rights per person, separate conversations, two-factor authentication, a shared memory that does not mix up each person's preferences. Nous Research offers part of this in its Hermes Business offer; Diadems provides it in every Team subscription.

Hermes Agent or OpenClaw: which to choose

  • OpenClaw, if…

    you want a personal assistant on your own computer, with the widest ecosystem of skills and companion apps (macOS, iOS, Android), and you know how to sift what you install.

  • Hermes Agent, if…

    you want an agent that works on a server, learns from its repeated tasks, delegates to sub-agents and also answers by email; and you prefer Python to Node.js.

  • Neither on its own, if…

    the agent has to serve a whole team, with nobody administering a server. You then need a managed agent, which adds isolation, accounts, tested updates and support.

Why Diadems chose Hermes Agent

Diadems runs Hermes Agent, and only Hermes Agent, for three reasons: it is designed to live on a server rather than on a personal computer, its learning loop makes an agent better over the months, and its architecture accommodates a real long-term memory, isolated per agent.

  • One virtual machine per agent, at Scaleway in the Paris region, shut off from private networks: the boundary that Hermes's security policy calls for.
  • Skills under control: the agent keeps its learning loop, but no dependency downloads itself while it is working.
  • Releases tested on a pilot agent before they are applied, while the agent is idle.
  • A team layer: accounts, rights per section, two-factor authentication, shared or private conversations.
  • Public patches. When a capability is missing, Diadems adds a minimal patch on its public fork and offers it to the project.

Want to run OpenClaw? Diadems is not built for that: a VPS or a rented instance will suit you better. The offers in detail: hosting Hermes Agent.

Frequently asked questions: Hermes Agent or OpenClaw

What is the difference between Hermes Agent and OpenClaw?

OpenClaw is a personal assistant written in Node.js, designed to live on your own hardware and answer in your messaging apps. Hermes Agent, from Nous Research, is written in Python and stands out for its learning loop: it writes and improves its own skills. Both are open source, under the MIT licence.

Is OpenClaw the same as Clawdbot or Moltbot?

Yes, it is the same project. Created by Peter Steinberger in November 2025 under the name Clawdbot, it became Moltbot at Anthropic's request, then OpenClaw at the end of January 2026. It is now run by the OpenClaw Foundation.

Hermes Agent or OpenClaw: which is more secure?

Both projects fixed vulnerabilities in 2026. OpenClaw also saw hundreds of malicious skills published on its ClawHub registry, which now scans every submission, and thousands of installations left reachable by their owners. Hermes has had no incident on that scale, but its own policy points out that the only real boundary is the operating system. Either way, isolate the machine that runs the agent.

Can you move from OpenClaw to Hermes Agent?

Yes. Hermes provides the hermes import-agent command, which takes over the settings, memory and skills of an OpenClaw agent. Secrets are only carried over on explicit request.

Hermes Agent or OpenClaw for a business?

Neither is designed to be shared by a team: both assume a single trusted user. For a business, you have to add accounts, rights per person, isolation and updates, either yourself or with a managed agent such as Diadems, which runs Hermes Agent in France.

Does Diadems host OpenClaw?

No. Diadems runs only Hermes Agent, which it tests and updates itself. For OpenClaw, a VPS or a rented instance is a better fit.

Sources, consulted on 10/11/2026: the github.com/openclaw/openclaw and github.com/NousResearch/hermes-agent repositories (README, releases, SECURITY.md); docs.openclaw.ai; the Hermes Agent documentation; the “Introducing OpenClaw” post (29/01/2026); Koi Security, “ClawHavoc” (February 2026); Censys and SecurityScorecard (January and February 2026); portal.nousresearch.com. OpenClaw and Hermes Agent belong to their authors; Diadems contributes to Hermes Agent without being affiliated with Nous Research.

Updated on .

Give it its first task.

Your agent is ready in a few minutes. Or tell us what you need first: we will help you choose its first assignment.

Create my agent