OpenClaw vs Hermes: what still differs, and how to choose
Both now write their own skills and remember in Markdown. Who makes each, their defaults and security records, memory, moving between them, and where each runs.
OpenClaw and Hermes Agent are now more alike than different. Both are free, MIT-licensed agents you run on a computer you choose; both remember you in plain Markdown files, write and refine their own skills in the same SKILL.md format, run jobs on a schedule, work with almost any model, local ones included, and talk to you from your chat apps or a desktop app. What still separates them: who is behind each (OpenClaw is run by a non-profit foundation with no paid tier; Hermes is made by Nous Research, a funded company that sells a hosted version and model plans), how each is guarded out of the box (OpenClaw trusts its single owner and leaves sandboxing and command approvals off; Hermes has an AI reviewer approve or refuse risky commands by default), and how each remembers. Each can import the other’s setup, so the choice is less permanent than it looks.
What they now share
The old line between them, Hermes the agent that writes its own skills against OpenClaw the agent with a marketplace, has gone. OpenClaw’s self-learning is on by default: the agent saves a procedure after hard multi-step work, patches a learned skill that misled it, and archives “skills nobody uses for 30 days.” Hermes’s agent creates, updates and deletes its own skills with a tool its docs call “the agent’s procedural memory,” and installs other people’s from skills.sh, LobeHub and ClawHub, OpenClaw’s own registry. Both follow the Agent Skills format, so a skill folder moves between them.
The rest of the overlap is as wide. Both speak MCP for tools and A2A for messaging other agents (OpenClaw’s A2A docs use a peer named “hermes” as their example). Both keep a SOUL.md, an AGENTS.md, a MEMORY.md and a USER.md. Both run scheduled jobs, connect to Telegram, WhatsApp, Discord, Slack, Signal, iMessage and more, and come as desktop apps for macOS, Windows and Linux. And both say the same thing about safety in different words: the agent acts with the permissions of the machine it runs on, so the machine is the boundary.
Side by side
As each project’s own pages stated it on October 9, 2026:
| OpenClaw | Hermes Agent | |
|---|---|---|
Made by | The OpenClaw Foundation, a 501(c)(3) non-profit formally introduced in July 2026; “no paid tier, hosted service, or token” | Nous Research, a company that announced $90 million in funding in October 2026 and sells model plans and a hosted Hermes |
Licence | MIT | MIT |
Built around | The Gateway, one always-on process that owns sessions, memory and every chat connection | One agent with several front ends sharing its state: a terminal interface, Hermes Desktop, a browser dashboard and a messaging gateway |
Current release | Weekly or faster; see the latest version | v0.21.6, October 8, 2026 |
Chat apps | 33 entries in its channel catalog | 28 platforms in its messaging docs |
Memory | Markdown files that grow, consolidated overnight by “dreaming,” on by default | Two capped files (about 800 and 500 tokens) plus searchable past conversations |
Own skills | Written and archived by the agent, on by default; each change can be undone | Written freely by the agent by default; a curator archives stale ones |
Risky commands, by default | Run without asking for the single trusted owner; sandboxing and approvals are opt-in | An AI reviewer approves low-risk commands, refuses dangerous ones and asks you about the rest |
GitHub stars | 391,534 | 252,201 |
Hosted by its makers | No | Yes: Hermes Cloud, from $0.51 a day while running (read October 11), model use extra |
Channel counts are not directly comparable: each project counts its own way, and OpenClaw’s catalog includes protocols such as A2A beside chat apps. What is useful is whether your app is on the list, and for the common ones it is on both.
Where they really differ
Who answers to whom
OpenClaw is stewarded by an independent foundation that employs the core team; its README says OpenAI is “a donor, not an owner” and that the project has no paid tier. Hermes is a product of Nous Research, whose October note announces $90 million raised and plans for “Hermes for Businesses,” and whose docs now call its Nous Portal subscription “the recommended way to run Hermes Agent.” Neither fact makes one agent better. They tell you what each project is paid to care about. OpenClaw’s docs make that argument at length in their own comparison with Hermes; Nous’s docs mention OpenClaw only as something to migrate from.
Defaults and the security record
OpenClaw’s docs are blunt: “Sandboxing and exec approvals are off by default. Default OpenClaw is a trusted single-operator assistant.” Hardening is something you configure, and its security audit tells you when you have drifted from it. Hermes starts guarded: its smart approval mode is the default, unattended jobs default to refusing anything that needs approval, an always-on blocklist refuses the most destructive commands whatever you set, and an unanswered prompt denies after five minutes. But Nous’s own security policy is as clear as OpenClaw’s that none of this is containment: “The only security boundary against an adversarial LLM is the operating system.”
The records are kept differently, which makes them hard to set side by side. OpenClaw publishes advisories on GitHub: 722 by October 9, 2026, 14 of them critical, the last critical one published in April, the latest batch of 75 on September 11; a Trail of Bits audit it published in September found no critical issues and two high ones. OpenClaw’s own caveat applies: those counts “describe disclosure records, not the number of currently exploitable defects.” Hermes’s own repository publishes no security advisories and lists its fixes in release notes, though GitHub’s advisory database lists ten CVEs against the hermes-agent package, published between May 26 and June 17, 2026. Its docs record a June 2026 campaign in which scanners reached Hermes dashboards left open to the internet and had the agent plant an SSH backdoor, after which sign-in became mandatory on any non-local address, and its October 8 release fixes dashboard sign-in flaws reported by Tenable Research. The lesson both records teach is the same: never put either agent’s control surface on the open internet.
How each remembers
OpenClaw writes what it learns into plain files in its workspace and keeps adding to them; a background pass it calls dreaming consolidates them, and search runs over an index of everything. Hermes keeps its working memory deliberately small: two files capped at 2,200 and 1,375 characters, written into the prompt at the start of each session, with every past conversation stored in a database the agent searches when it needs more. Both are readable and editable; the difference is that OpenClaw remembers more by default and Hermes curates harder.
One agent or several
OpenClaw can run several independent agents from one Gateway and bind each to particular chats, even on one WhatsApp number. Hermes runs one agent across your surfaces and uses profiles for separate instances, and its FAQ notes each profile needs its own WhatsApp number; for teams of agents it has Bot Mode and a kanban board that profiles work from.
Moving from one to the other
Each project ships an importer for the other. OpenClaw’s is openclaw migrate hermes (and “Import from Hermes” in its setup); Hermes’s is hermes claw migrate, which its setup offers when it finds ~/.openclaw. Both preview the changes first, and both bring across the persona file, memory, skills, the model and provider settings and MCP servers. Neither brings conversation history, scheduled jobs or plugins as working state: those are archived for you to set up again, and WhatsApp needs pairing again. API keys move only if you say so, and OpenClaw’s migration guide warns against leaving both agents signed in with the same refreshed login. Run them side by side for a while if you like, but not on the same chat accounts, for the reason any two copies of one agent fight over them.
Where each one runs
Both are meant to run around the clock on a computer that stays on: a spare laptop, a Mac mini, a home server or a cloud server. OpenClaw’s project offers no hosted version, so the alternative to running it yourself is a host: managed OpenClaw hosting is a field of its own. Nous sells Hermes Cloud, priced on its own page on October 11, 2026, at $0.51 a day for a dedicated instance while running (2 GB of memory, 2 vCPUs) or $0.99 for the larger size, $0.03 a day stopped, with model and tool use billed on top from your Nous credit. Where to run Hermes sets it beside the other places to run it.
On Everpod, OpenClaw comes both ways and Hermes one. The OpenClaw pod puts your agent on its own private cloud computer, set up for you with no technical knowledge needed; your pod arrives secured, backed up, and private to you. $29 a month, cancel anytime. Your plan covers up to $10 of model usage a month. Hermes, or OpenClaw if you would rather run it yourself, comes installed on a developer pod, which you run yourself: a permanent cloud computer (VPS), reached only through your own private network, from $24 a month, on your own plan or API key.
How to choose
Choose OpenClaw if you want the larger project, with the most integrations and the most answers when you search an error, several agents bound to different chats, and a project with no company behind it, and you are content to switch on its guards yourself where you want them. Choose Hermes if you would rather start with approvals on, like a small curated memory and a desktop app built around the agent, or want its makers to host it for you. Either way, give it a machine of its own and keep its control panel off the open internet.
Dating note: both projects ship every week or two; every figure here was read from their own pages and repositories on October 9, 2026.