What is Hermes?

The open-source agent harness from Nous Research. Mobile-first. Multi-model. The main alternative to the OpenCLAW harness I run.

The first time a CEO peer asked me about Hermes was a few weeks ago. The second time was a last week. Now, it’s coming up multiple times per day. So… I felt like it’s time to write about it.

I have not installed it. I run [OpenCLAW][1] which is the [harness][6] I know inside out. But Hermes is showing up in enough conversations that ignoring it would be a mistake.

So here is what I have gathered, from the Nous docs, from the people running it, and from the public discourse. The honest version, with the parts I don’t know much about flagged.

What Hermes is

Hermes is Nous Research's open-source agent harness. Standalone, multi-model, and has a Telegram-first interface.

Nous is the same group that ships the Hermes family of large language models, but this is a different product: a harness. MIT-licensed. The source lives at [github.com/NousResearch/hermes-agent][2]. It has around 170,000 GitHub stars as of late May 2026.

Here’s a simple description of what it is: persistent memory via a local [SQLite database][8] with full-text search (so the agent can recall context across sessions without re-pasting). A [skills][3] system that follows the same open agentskills.io [SKILL.md][4] standard as OpenCLAW and Claude Code, which means a skill folder you wrote for one of them works on the others. A natural-language [cron][5] scheduler. Subagent delegation for parallel work. A multi-platform gateway: Telegram, Discord, Slack, WhatsApp, Signal, Email, CLI. Model-agnostic backend: works with Anthropic, OpenAI, xAI's Grok via X Premium credits, and OpenRouter.

The thing that stands out: Telegram is the headline interface. The picture Nous paints in their docs and demos is the operator running their company from their phone, with the harness wired to the LLM and the connectors in the cloud. That is a different operating model from a laptop-resident terminal session. Most operators run Hermes on a [cloud VM][9] so Telegram has a 24/7 endpoint to talk to.

A different shape of harness

The reason a CEO should know what Hermes is, even if they end up not running it, is that the existence of Hermes proves the harness market has split into two operating models.

Mobile-first changes where the work happens. If your harness lives in a terminal on your laptop, your harness is at your desk. If your harness lives in Telegram on your phone, your harness is wherever you are. Some CEOs find that the unlock; some find it the wrong abstraction. For an operator who runs most of their workday from their phone (between meetings, in the back of an Uber, on weekend mornings before the kids wake up), Hermes is a meaningfully different lifestyle than the laptop pattern. The case for the laptop-terminal pattern, and why most CEOs should start there before they reach for a harness at all, is [why CEOs should use Claude Code in the terminal][10].

Model-agnostic is a hedge. Hermes was built on the assumption that the model layer would commoditize and that operators would want to keep the loop while swapping the engine. OpenCLAW also works with multiple models, but Hermes leans harder into "swap the backend, keep the loop." For a CEO who wants to avoid lock-in to any single provider (including Anthropic), Hermes is attractive.

The xAI integration is the headline of 2026. In May, Nous shipped a Hermes-to-Grok integration that lets users use their X Premium Grok credits as the model backend instead of paying for API tokens. For an operator already paying for X Premium, the marginal cost of running an agent through Grok approaches zero. The cost story is doing a lot of the work in why Hermes is having a moment right now.

I have not switched. The OpenCLAW ecosystem still earns its keep for the shape of work I do, and switching costs are real.

What good Hermes usage looks like

Honest caveat first: I have not run Hermes long enough to write the operator's-eye view from the inside. What follows is what operators who do run it report, distilled from X and from CEO peers who have been on it for at least a quarter.

The lean-setup pattern. Operators describe Hermes as leaner and more stable than alternatives. The implication is a smaller surface area to configure, fewer ecosystem packages to learn, and less time getting it to production-grade. For a CEO who values a smaller stack they can hold in their head, this is the strongest argument for Hermes.

The phone-first workflow. A pattern that repeats across X: a CEO sets up Hermes on a cloud VM, wires it to Telegram, and then runs their day through DMs to the agent. "What's on the calendar today?" "Draft the follow-up for the 11 AM call." "What's slipping?" The agent answers in Telegram in seconds, with the full context of the connectors behind it. The on-laptop terminal pattern is built into the OpenCLAW workflow; the in-phone Telegram pattern is built into the Hermes one.

The Grok-backed agent. With the May 2026 xAI integration, Hermes users who have X Premium can run agents off Grok credits. Lower marginal cost than paying API rates at scale. The trade-off is model quality, which is the next section.

The ecosystem caveat. Hermes is younger than OpenCLAW. The community library of skills, connectors, and workflows is smaller. Operators who want to lift-and-shift a stack are more likely to find what they need in the OpenCLAW ecosystem; operators willing to write their own skills will be fine on either.

The supervisor-and-specialists pattern (one agent that delegates to a small fleet, made famous by JP Morgan's internal "Ask David" architecture) gets implemented on top of both Hermes and OpenCLAW. The harness layer abstracts away the specifics; the architecture pattern is portable.

Common mistakes

The anti-patterns visible from the discourse on X and from CEO conversations:

Treating Hermes as a chat client. The Telegram interface is friendly enough that some operators never write a skill. They just chat with the agent. That works for two weeks; without skills, the leverage stops compounding. A harness without a skill catalog is a chatbot with extra steps.

Wiring xAI before testing Anthropic. The Grok integration is a recent headline, and the cost savings might matter. But for CEO-grade tasks (writing, analysis, judgment, long-context synthesis), Anthropic models still produce stronger output than Grok 4.x. Use Grok where the cost story matters most: high-volume, low-stakes, repetitive work. Use Claude where the quality story matters: anything that goes to a customer, a board, or a court. I do, however want to shout out Grok’s superior knowledge of current events - it’s always great if you want to know what’s happened over the past 7 days or so.

Underestimating the cloud-VM step. OpenCLAW can run locally on a laptop you own; you install it on your machine and start. Hermes can also run laptop-local, but the Telegram-first workflow only really pays off if the agent has a 24/7 spot to receive your DMs from anywhere. That means a cloud VM, which means a small setup project: renting a box from a provider like Linode or DigitalOcean, getting it secured, and having a backup plan for when something drifts. The cloud-VM explainer linked above walks through the parts. Not hard, but a step new operators sometimes hit as unexpected friction.

Expecting parity with OpenCLAW on day one. Hermes is younger. Some features the OpenCLAW ecosystem has had for months are still maturing in Hermes. Don't expect a like-for-like swap; expect a different operating model with different trade-offs.

Where Hermes stands today

Active development under Nous Research. Around 170,000 GitHub stars in late May 2026; the commit cadence shows no sign of slowing. Skills portability via [agentskills.io][4] means the skill catalog you build on either harness travels with you if you switch.

The open-source agent-harness market is effectively a two-horse race: Hermes and OpenCLAW. Other harnesses exist, but neither matches the GitHub star count or the active operator community of these two. The two-horse race is the shape of the harness market in 2026.

Do this next

If you are choosing between Hermes and OpenCLAW, read the side-by-side: [OpenCLAW vs Hermes][7]. If you are already running Hermes, I’d love to hear about your experience.

[1]: /articles/what-is-openclaw [2]: https://github.com/NousResearch/hermes-agent [3]: /articles/what-are-skills-in-claude-code [4]: https://agentskills.io [5]: /articles/what-is-a-cron [6]: /articles/what-is-a-harness [7]: /articles/openclaw-vs-hermes [8]: /articles/what-is-a-sqlite-database [9]: /articles/what-is-a-cloud-vm [10]: /blog/why-ceos-claude-code-terminal