TL;DR
Nearly every comparison of these two ranks them by GitHub stars and concludes OpenClaw wins. That measures adoption velocity, not fit. The repositories show a cleaner split. OpenClaw spends its engineering on reach: alongside its TypeScript core it ships native Swift and Kotlin clients, which is why "Any OS. Any Platform." is in its own description. Hermes Agent spends its engineering on accumulation and containment: memory that compounds into auto-generated skills, plus five sandbox backends with container hardening and namespace isolation.
Pick on that axis. If you need the agent everywhere you already are, OpenClaw. If you need it to get better over time and stay in a box while it does, Hermes.
Key Takeaways
- Star count is the wrong signal. OpenClaw's repo is four months younger than Hermes' and has roughly 158,000 more stars. That gap measures virality.
- OpenClaw ships native mobile. Swift and Kotlin sit in its language breakdown. Hermes has neither.
- Hermes treats isolation as product surface, offering local, Docker, SSH, Singularity and Modal backends with namespace isolation.
- Both auto-generate skills from experience, so "it writes its own skills" does not separate them.
- Both are MIT. GitHub flags OpenClaw's licence as "Other" only because of a trailing pointer to third-party notices.
- Language matters if you will extend it. Hermes is Python-dominant, OpenClaw is TypeScript-dominant.
- Neither one solves token cost or context discipline. Those remain yours.
The comparison everyone repeats
Search either name and you get the same paragraph. Both are open-source, self-hosted personal AI agents. Both remember things across sessions. Both plug into your chat apps. Both run shell commands and drive a browser. OpenClaw has more stars, so OpenClaw.
Every sentence there is true and the conclusion does not follow.
Two other things worth flagging, because the secondary coverage got them wrong and it is easy to check. One widely-shared guide credits Hermes with "16+ messaging platforms". Nous Research's own site lists Telegram, Discord, Slack, WhatsApp, Signal, Email and CLI. Another cites star counts that were already months stale when published. If a comparison cannot get the platform list right, its recommendation is not worth much either.
What the repositories actually show
Figures pulled from the GitHub API on 7 August 2026. Star counts move; the structural facts underneath them do not.
| Hermes Agent | OpenClaw | |
|---|---|---|
| Publisher | Nous Research | OpenClaw Foundation |
| Repo created | 22 July 2025 | 24 November 2025 |
| Stars | 226,989 | 385,461 |
| Licence | MIT | MIT (+ third-party notices) |
| Primary language | Python | TypeScript |
| Native mobile code | none | Swift, Kotlin |
| Self-description | "The agent that grows with you" | "Any OS. Any Platform." |
Those last three rows are the whole article. One project's code is spread across native mobile runtimes. The other's is concentrated in one language with reproducible-deployment tooling around it. Both descriptions are accurate advertisements for where the effort went.
Where OpenClaw goes wide
OpenClaw's design goal is to be wherever you already are, and it spends real engineering to get there.
It reaches you through WhatsApp, Telegram, Slack, Discord, Signal and iMessage, which is a longer surface than Hermes covers, and iMessage in particular is not a trivial integration. The native Swift and Kotlin clients mean it is not merely reachable from a phone but running on one. Its memory and skills are plain Markdown and YAML on your disk, which makes the whole state inspectable with a text editor and keeps the extension format low-friction enough for a large community to publish into.
The result is the fastest-growing repository of its kind, which is exactly what breadth optimised for distribution produces.
The cost of that choice is surface area. Every platform is a maintenance obligation, and an agent that runs shell commands, drives your browser, reads your files and sends your email is holding a lot of authority in a place designed for convenience.
Where Hermes goes deep
Hermes is trying to become more useful over time rather than more reachable.
Its own framing is "the agent that grows with you", and the mechanism is specific: it learns your projects, auto-generates skills when it solves something, and does not forget how it solved it. Memory lives in ~/.hermes/ with no cloud dependency. On top of that sit natural-language scheduling, subagent delegation for parallel work, and web browsing.
The part that gets least attention and matters most is sandboxing. Hermes offers five execution backends: local, Docker, SSH, Singularity and Modal, with container hardening and namespace isolation. Singularity is an HPC runtime and Modal is serverless, which tells you the intended range runs from a laptop to a research cluster.
It is also deliberately model-neutral, working through Nous Portal, OpenRouter, any OpenAI-compatible endpoint, or a local vLLM deployment. If you care about not being tied to one provider, that is depth in the direction you want. Our self-hosting vs API breakdown covers that trade in general.
The question nobody asks: how contained is it?
This is the part worth slowing down on, because it is where the two genuinely diverge and it barely appears in any published comparison.
Both of these agents execute shell commands, read and write your files, and drive a browser with your sessions in it. That is the entire point, and it is also a standing grant of authority to a probabilistic system. The relevant question is not "can it do things" but "what is the blast radius when it does the wrong thing".
OpenClaw's answer is local-first: it runs on your machine, your data stays there, and you inspect state as Markdown on disk. That is a coherent and defensible position, and for a personal assistant it is often the right one.
Hermes' answer is that execution isolation is a first-class product surface, with five backends and namespace isolation available before you have written a line of configuration. That is a different posture, and it is the posture you want if the agent is going to run unattended on a server, on a schedule, spawning subagents while you sleep.
Scheduled autonomy and containment are the same conversation. An agent that only acts when you message it has a natural throttle. One that wakes on cron does not, and at that point the sandbox is the throttle.
Why the star gap is misleading
OpenClaw has roughly 158,000 more stars than Hermes. It also created its repository four months later.
Star velocity like that measures how shareable a project is. OpenClaw is unusually shareable: it has a mascot, a memorable rename story, it lands in consumer chat apps, and the demo is legible to anyone who has ever sent a text message. Hermes is a Python daemon from a research lab whose headline feature is that it slowly becomes more useful. One of those spreads on social media and it is not the one measuring depth.
Use stars to gauge how quickly bugs get found and how many community skills exist. Do not use them to decide which architecture matches your constraint.
Which one to pick
Choose OpenClaw if you want a personal assistant reachable from every device and chat app you use, you are comfortable with it holding broad authority on your own machine, you would rather edit Markdown than configuration, and TypeScript is the language you would extend it in.
Choose Hermes Agent if the agent will run unattended on a server, isolation is a requirement rather than a preference, you want capability that compounds across months instead of resetting, you need model neutrality including a local vLLM endpoint, and Python is your extension language.
Run both if you are honest about the fact that they are not really the same product. A reachable assistant on your phone and a contained worker on your server are two jobs.
What neither of them solves
Both are agent runtimes, not economics. Both will happily spend tokens on your behalf, and an agent with memory and cron spends them while you are not watching. Understanding where that goes is a separate discipline, covered in AI agent costs.
Both also inherit every context problem the underlying model has. Persistent memory makes this sharper rather than softer, because accumulated context is still context and it can still rot or poison. See agent memory for how retention strategies differ, context engineering for managing what actually reaches the model, and agent skills for what auto-generated skills are and where they break.
Frequently asked questions
What is the main difference between Hermes Agent and OpenClaw?
OpenClaw optimises for reach, shipping native Swift and Kotlin clients so it runs on more platforms, including iMessage. Hermes Agent optimises for depth: memory that compounds into auto-generated skills, subagent delegation, and five execution sandbox backends with namespace isolation. Both are MIT-licensed and self-hosted.
Is OpenClaw better than Hermes Agent because it has more stars?
No. OpenClaw has roughly 158,000 more stars despite creating its repository four months later, which measures shareability rather than fit. OpenClaw is easy to demo in a chat app and has a strong viral story. Hermes is a Python daemon whose main feature accrues over months. Choose on architecture, not popularity.
Which is better for running an agent unattended on a server?
Hermes Agent, primarily because of sandboxing. It offers local, Docker, SSH, Singularity and Modal backends with container hardening and namespace isolation, and pairs that with natural-language scheduling and subagents. Scheduled autonomy removes the natural throttle of a human sending a message, which makes containment the thing that limits blast radius.
Are both Hermes Agent and OpenClaw really open source?
Yes, both under MIT. GitHub's licence detector reports OpenClaw as "Other" only because its LICENSE file appends a line pointing to THIRD_PARTY_NOTICES.md after the standard MIT text. The grant itself is unmodified MIT.
Which models do they support?
Hermes is explicitly model-neutral, working through Nous Portal, OpenRouter, any OpenAI-compatible endpoint, or a local vLLM deployment. OpenClaw is bring-your-own-key and runs against cloud or local models. Neither locks you to a single provider.
Do both write their own skills?
Yes, and this is why it does not separate them. Hermes auto-generates a skill when it solves a problem and stores it for reuse. OpenClaw uses a portable skill format that the community publishes into heavily. The difference is provenance: one grows skills from your own usage, the other imports them from a large ecosystem.
Sources
- NousResearch/hermes-agent: repository metadata, licence and language composition
- openclaw/openclaw: repository metadata, LICENSE file and language composition
- Hermes Agent, Nous Research: capabilities, sandbox backends, memory and platform list
- OpenClaw: official project site
Repository figures were read from the GitHub API on 7 August 2026. Star counts change daily; the licence, language and architecture facts are the durable part.
