Comparison · checked October 17, 2026

Hermes vs OpenClaw. Two open source agents, two philosophies.

Hermes Agent and OpenClaw are the two most widely used open source AI agents for putting an AI to work inside your tools and messaging apps. Both are MIT-licensed, run on your own infrastructure and learn from their tasks. They differ in architecture, default security, memory and target audience. This comparison keeps only what can be verified in their official documentation.

Book 20 minutes with us See the comparison

Hermes Agent is an open source project by Nous Research. Kailli is not affiliated with Nous Research.

Updated October 17, 2026

01In short

The difference in three sentences

If you only have a minute, remember this.

Hermes Agent: a work agent

Written in Python by Nous Research, it is built for work first: an agent that carries out tasks, isolates its commands in Docker or on another machine, and builds up know-how by writing its own skills.

OpenClaw: a connected personal assistant

Written in TypeScript and backed by a foundation, it is built as a personal assistant first: a gateway running on your machine, apps for computer, phone and watch, and the widest range of messaging apps.

What they share: it all depends on the setup

Both genuinely act inside your tools. Their real-world security depends less on the software than on the configuration: who can talk to them, what they can see, where their commands run.

02The table

Hermes Agent and OpenClaw, point by point

Sources: GitHub repositories and official documentation of both projects, consulted on October 17, 2026.

Hermes Agent and OpenClaw compared
Criterion	Hermes Agent	OpenClaw
Publisher	Nous Research, an AI research lab	OpenClaw Foundation, a nonprofit organization
License	MIT	MIT
Language	Python	TypeScript, on Node.js
Repository created	July 2025	November 2025
Positioning	a work agent that learns: memory and skills built through use	a personal assistant on your devices and messaging apps
Installation	installer for Linux, macOS, WSL2 and Windows, desktop app, Docker, Android	installer for Linux, macOS and Windows, Windows app, npm, Docker
Interfaces	terminal, desktop app, messaging apps, code editors	web dashboard, terminal, messaging apps, macOS, iOS, Android and watch apps
Messaging	over 20 platforms, including Discord, Slack, Teams, Telegram, WhatsApp, email	over 20 channels, including iMessage, WhatsApp, Signal, Teams, Slack, Discord
Memory	short notes always loaded, plus search across the full session history	Markdown files (profile, memory, daily notes) automatically consolidated in the background
Skills	created and improved by the agent, open agentskills.io standard, Skills Hub catalog	learned by the agent with history and undo, public ClawHub catalog
Scheduled tasks	built-in scheduler, result delivered to your messaging app	scheduled automations, webhooks, email triggers
Where commands run	your choice: local, Docker, SSH, or remote sandbox (Modal, Daytona…)	on the host machine by default; Docker, Podman or SSH sandbox to be enabled
Human approval	dangerous commands require approval; no answer means denied by default	adjustable execution policies, from "deny all" to "allow all," with approval prompts
Default access	list of authorized users per messaging app	unknown senders get a pairing code; groups on an allowlist
Multiple agents	separate profiles, each with its own memory and settings	several isolated agents in one gateway, routed by account or channel
External tools (MCP)	yes	yes
Migration	official import of an OpenClaw configuration	memory import from Hermes, Claude Code or Codex

03In detail

The real differences, topic by topic

The table says what; here is what it changes day to day.

Memory

Hermes keeps two short, deliberately limited notebooks (your preferences, its working notes), loaded at every session, and can search its entire conversation history. External memory backends can be plugged in. OpenClaw writes Markdown files: a profile, a long-term memory and daily notes, then a background consolidation process surfaces what matters. In both cases, memory can be read and corrected like a document.

Skills

Both agents learn procedures and write them down as cards. Hermes follows the open agentskills.io standard, which makes its cards reusable by other compatible agents. OpenClaw versions learned cards, lets you undo a change and archives those unused for 30 days; its ClawHub catalog shows a security audit before installation.

Where commands run

This is the most important difference for a business. Hermes lets you run its commands somewhere other than the machine hosting it: a Docker container, a remote server over SSH or a cloud sandbox. OpenClaw keeps the gateway on the host machine and offers a sandbox that is off by default, which its own documentation describes as a useful limit but not a perfect security boundary.

Who can talk to it

OpenClaw starts from cautious settings: gateway reachable only locally, a pairing code for any unknown sender, groups on an allowlist and a configuration audit command. Hermes works with an explicit list of authorized users per messaging app and an approval system for dangerous commands, which can be manual or assisted by a second model.

Devices and messaging apps

OpenClaw's advantage: apps for macOS, iOS, Android and watch, access to the camera, screen and notifications, iMessage support. Hermes stays focused on computers and servers, with a desktop app, the terminal, code editors and team messaging apps.

Teamwork

OpenClaw is designed around one gateway per circle of trust: one person or a team whose members trust each other, not users who are wary of one another. To separate uses, you need separate gateways. Hermes separates uses through profiles, each with its own memory, skills and access; two teams that must share nothing each get their own profile, ideally on separate accounts.

Installation and administration

Both install in a few minutes with an official script. OpenClaw can reuse an existing Claude Code or Codex login and opens a web dashboard. Hermes is configured in the terminal or from its desktop app. In both cases, the exact commands change from one release to the next: follow the official documentation.

04Which one to choose

Hermes Agent or OpenClaw: it depends on your needs

There is no universal winner. The right choice depends on what the agent must do, who runs it and how much isolation you expect.

Lean toward Hermes Agent

For a work agent installed on a company server: scheduled tasks, research, writing, technical operations, with commands isolated in Docker and shareable skills. It's the one we use every day at Kailli.

Lean toward OpenClaw

For a personal assistant that is everywhere: on the phone, the watch and the computer, in iMessage or WhatsApp, with a web dashboard and a TypeScript ecosystem.

Either one works

For a simple first use (weekly summary, news watch, message drafts) in a small team that trusts each other. In that case, pick the one your technical team will be able to maintain.

Switching from one to the other

Hermes Agent provides an official tool to import an OpenClaw configuration (memories, skills, settings), with a preview before any change. OpenClaw can import a Hermes agent's memory. So the choice is not final.

05Before you deploy

Five rules, whichever agent you pick

These are the same rules we apply in-house and for our clients.

A dedicated account for the agent, never that of an executive or an administrator.
Commands run in a container or on a separate machine, not on the workstation.
A written list of people authorized to talk to it, reviewed regularly.
Human approval on anything that sends, pays or deletes.
Skills installed from a catalog only after reading them: a card can contain malicious instructions.

The details of the security framework on the Hermes Agent side are in our Hermes Agent installation guide.

06Our method

How this comparison was made

An honest comparison is dated and sourced.

Only facts read in the official repositories and documentation of both projects.
No volatile figures (stars, downloads): they change every day.
No disparagement: both projects are active, serious and useful.
We use Hermes Agent every day: we say so, so you can take it into account.
Next review at every major release of either one.

Before choosing, look at concrete AI agent use cases in a business setting.

Torn between the two? We'll look at your needs in 20 minutes.

Book 20 minutes

Written by Rémi Carbonnefounder and CTO of Kailli

Frequently asked questions: Hermes vs OpenClaw

What is the difference between Hermes Agent and OpenClaw?

Hermes Agent, by Nous Research, is a work agent written in Python: memory, skills built through use and commands isolated in Docker or on another machine. OpenClaw, backed by a foundation, is a personal assistant written in TypeScript, available on computer, phone and watch. Both are open source under the MIT license.

Are Hermes Agent and OpenClaw free?

Yes, both are open source software under the MIT license. You pay for the AI model you use and, where applicable, hosting.

Which one is more secure?

Neither is secure by magic. OpenClaw starts from cautious access settings (pairing, allowlists, audit); Hermes makes isolated command execution and approval of dangerous actions easier. Real-world security depends on your configuration: permissions, secrets, isolation and authorized people.

Which one should a business choose?

For a server-based work agent, shared by a team and governed by written rules, we recommend Hermes Agent. For a personal assistant on one person's devices, OpenClaw is very comprehensive.

Can they use the same models?

Yes. Both work with the major providers (Anthropic, OpenAI, Google, OpenRouter) and with local models.

Can you migrate from OpenClaw to Hermes Agent?

Yes. Hermes Agent offers an official import of an OpenClaw configuration (memories, skills, settings), with a preview before any change.

Twenty minutes. And you'll know whether we can help.

No sales pitch. You tell us where you stand with AI, we tell you how far you can go.

Book 20 minutes Prepare for the call