OpenClaw vs Hermes vs Grok Bot: what's actually different, and where each one runs
Two open-source harnesses you host and one closed product run for you in the cloud. Licence, where the machine is, what persists, model choice, memory, channels and price in one table, with the hosting column the popular comparisons skip.
All three give you a persistent agent with a computer that keeps its files and logins between sessions. The difference is who owns what. OpenClaw and Hermes are MIT-licensed software you run on a machine you choose (a laptop, a server, a managed pod), with any model behind them and their state in files you can read. Grok Bot is a closed product on a cloud computer run for you, one per account shared by all your bots, with the model chosen for you, included with xAI’s The X Subscription from $30 a month. The open pair are now more alike than different: both live in your chat apps and write their own skills, and what separates them is who is behind each and how each is guarded out of the box. Pick the closed one for zero setup, and one of the open two for an agent that is yours.
The table
| OpenClaw | Hermes | Grok Bot | |
|---|---|---|---|
Licence | MIT, open source | MIT, open source | Closed, proprietary |
Where it runs | Any machine you choose: laptop, VPS, managed pod | Any machine; Nous also runs Hermes Cloud | Cursor’s cloud only |
Machine model | Your machine; one or many agents per Gateway | Your machine; named bots in Bot Mode | One shared computer per account for every bot |
What persists | Everything on the disk you gave it | Everything on the disk you gave it | /workspace, browser state and sign-ins; installed packages are replaceable |
Models | Any provider, including local models, or a ChatGPT subscription | Any provider, including local models | Chosen for you, no picker; Elon Musk named Claude Opus 5.5 among the models it will use (October 7, 2026) |
Memory | Markdown files plus SQLite; overnight dreaming; self-written skills | Files plus SQLite; self-written skills, curated | Per-bot, undisclosed, no file export |
Where you talk to it | Telegram, WhatsApp, Discord, Slack and the rest of its channel catalog; a web dashboard; its own apps: macOS, Windows, Linux, iPhone, Android | 28 platforms in its messaging docs; a desktop app | Its own apps: macOS, Windows, Linux, iPhone, iPad, Android; a team’s shared bot also in Slack |
Software price | Free | Free | Bundled with The X Subscription (from $30/mo); weekly usage, then prepaid extra usage |
The hosting column is the one the popular comparisons skip
The viral version of this comparison had rows for licence, models and memory and said “self-hosted” for the open pair as if that settled where they run. It does not. “Self-hosted” spans a laptop that sleeps, a Mac mini in a cupboard, a $6 server you administer, and a managed machine someone else keeps patched. Each is a different product experience with the same software, and the choice between them is most of the decision for an open agent. Grok Bot removes the choice: the computer is in Cursor’s cloud, provisioned and updated for you, and every bot on your account shares it. Their FAQ is direct: “The computer is assigned per user, not per Bot,” and their security guidance follows, “Do not use separate Bots as a security boundary.”
The persistence row is where the word “computer” gets tested
On the open pair, persistence is whatever your disk holds: the agent’s installed tools, its keys, its browser profile and its notes all survive a restart because nothing rebuilds the machine under it (unless you run it in a container and mount too little, which is a known trap). Grok Bot’s docs draw a narrower line: files in /workspace, browser state and supported sign-ins “are designed to survive normal computer updates,” while temporary directories, manually installed packages and uncommitted state are “replaceable.” Users translated that into a habit: reinstall your tooling after updates and keep everything that matters in the one folder. For an agent that accumulates competence in the tools it sets up, that is a real difference in what “its own computer” means.
OpenClaw against Hermes, briefly
By October 2026 the two have grown alike: both write and prune their own skills, keep their memory in Markdown files and reach most chat apps, and the longer comparison covers what still separates them: who is behind each, how each is guarded by default, and how each remembers. Both run any model. Both now do multi-agent: OpenClaw with several agents on one Gateway and tools for spawning and messaging sessions, Hermes with a named roster in Bot Mode. Nous also sells Hermes Cloud, its own hosted version, billed by the day.
Grok Bot against both
It was the first frontier-lab version of the idea, since joined by OpenAI’s dots and Meta’s Muse, and the reasons to choose it are the reasons to choose any polished closed product: nothing to install, apps on every platform, a roster of teammates that message each other and come back for approvals, and a workflow you can teach by recording your screen. The reasons people look past it are equally concrete: the model is chosen for you, the computer is shared across your bots, the state cannot be exported as files, your own bots have no Telegram or WhatsApp channel (you talk to them in Grok Bot’s apps), and the price is a subscription bundle rather than a line you control. The first of those changed shape in October 2026: on October 7 Elon Musk wrote on X that Grok Bot will use “the best back end model for any given task, including Claude Opus 5.5, MidJourney, Suno and other leading APIs,” so a bot may run on another lab’s model, while xAI’s docs say there is no model picker and that the serving mix “can change over time with no fixed vendor set guaranteed.” The two-way comparison goes deeper on the ownership question.
How to choose
If you want teammates that appear this afternoon and you already pay for a plan that includes it, Grok Bot. If you want an agent that is yours to move, model and inspect, one of the open two: OpenClaw for the larger project with no company behind it, Hermes to start with approvals on or to have its makers host it. Then decide where it lives, because that decides whether you become a system administrator. A laptop is fine to try. A server you run is the cheapest always-on option and the most work. Managed OpenClaw hosting is the middle path; Everpod’s version is a private machine for your agent, never shared, at $29 a month with up to $10 of model usage included, and the agent still lives in your own messaging apps. Whatever you pick, read the persistence row twice. It is the one that decides what you still have in a year.