What is the Hermes agent runtime (and how it differs from the Hermes model)?

The Hermes agent runtime is open-source software from Nous Research (MIT-licensed) that turns a language model into a persistent agent. It handles memory, tools, channels, and model access; it is not itself a model. That distinction is the whole point of this page, because "Hermes" now names two different things from the same lab.

Hermes the runtime vs Hermes the model

Most explainers stop at "brain vs body," then never tie it to a decision you actually make. Here's the operator version.

The Hermes model family (Hermes 2, 3, 4) are weights — fine-tuned LLMs you'd serve behind an inference endpoint. They answer prompts. They don't remember yesterday, don't read your Telegram, and don't write their own tools.

The Hermes agent runtime is the software around a model. It maintains memory across sessions, wires up channels, exposes tools, and decides what to send the model on each turn. Swap the model underneath and the agent keeps its memory and its skills. That's what "model-agnostic" buys you: the runtime is the part that persists.

The two only meet in one honest place — through OpenRouter you can route the runtime to Nous's open-weight Hermes models. It's a legitimate combination, not a requirement. The runtime works just as well driving Claude, GPT, or Gemini.

Who builds the Hermes runtime, and is it open source?

It's built by Nous Research and released under the MIT license; Nous Research publishes it at hermes-agent.nousresearch.com. Being MIT-licensed means you can self-host it yourself — or run it as a managed instance on Lobsterland, on the same account, dashboard, and billing as OpenClaw. The runtime is the same software either way; only the operational burden changes.

What the Hermes runtime actually does

Independent of where it runs, the runtime's job is to stand up a single, persistent agent:

• Persistent memory across sessions, so it doesn't reset between conversations.
• A self-improving learning loop — after solving hard tasks it writes its own reusable skills, plus you can install skills from the Hermes Hub.
• Channel plumbing so people can reach the agent where they already are.
• Model access through whichever provider key you supply.

This is why Nous frames Hermes as "a different kind of agent, not a lighter OpenClaw" — it's memory-first by design.

What the Hermes runtime exposes on managed Lobsterland

This is the part every ranking page skips: when someone else runs the runtime for you, what surface do you actually get? On Lobsterland, a Hermes instance exposes a deliberately scoped set — and we'd rather name it precisely than imply the full self-host feature list ships managed.

• Chat — dashboard and in-browser chat with the agent.
• Channels — Telegram and Slack only. Allowlist a specific user, a whole channel, or anyone, and require an @mention in groups so the agent replies only when called.
• Logs — runtime logs to watch exactly what the agent does.
• Skills — the agent's own self-written skills plus Hermes Hub installs (the learning loop, not a heavy management console).
• Bring-your-own model keys — Anthropic, OpenAI, OpenRouter, Google, and MiniMax. Via OpenRouter, Nous's open-weight Hermes models.
• Subscription login, with one caveat — ChatGPT (Codex) OAuth works as advertised. Claude (Claude Code) OAuth is wired and authenticates, but Anthropic restricts Claude subscription use from third-party and hosted clients, so that traffic can consume additional metered usage rather than your Claude Pro/Max plan — for Anthropic, a plain API key is the predictable choice. Google's Gemini CLI OAuth isn't supported at all; use a Google API key.

It runs as an isolated managed instance with no public IP, no servers, Docker, or Kubernetes to maintain, with persistent memory across sessions — and it launches in about two minutes. Newer Hermes builds (v2026.6.5+) also add a native web dashboard, but treat that as version-dependent, not a baseline guarantee. For the model, Hermes takes either a plain provider key or a subscription login. ChatGPT (Codex) OAuth works as advertised. Claude (Claude Code) OAuth is wired and will authenticate, but Anthropic restricts Claude subscription use from third-party and hosted clients, so expect it to draw additional metered usage rather than your Claude plan — budget it as API spend, or use a plain Anthropic key. Google's Gemini CLI OAuth isn't supported; use a plain Google API key for Gemini.

The full capability matrix lives on the Hermes runtime hosting page if you want the side-by-side.

Where self-host (and OpenClaw) wins

Honesty matters more than the pitch here. The managed Lobsterland Hermes surface intentionally does not carry several things the broader OpenClaw runtime does:

• WhatsApp channels (Hermes channels are Telegram + Slack only)
• Workspace / file editor
• Cron and scheduled work
• Multi-agent routing
• Hosted Browser
• Usage analytics

If you need that fuller operating surface today, run OpenClaw instead — it's the broader runtime, and you can run both on one account. And if your reason for choosing Hermes is total control over every upstream feature, raw self-hosting the open-source runtime gives you the entire surface — at the cost of patching, gateway config, and backups you'd otherwise never touch.

What it costs to run a Hermes agent managed

Managed Hermes runs on the same capacity tiers as OpenClaw: Light $6.90/month, Medium $14.90/month, Large $29.90/month. Model usage is billed by your own provider with no markup — you pay the model vendor directly, and your keys stay encrypted on your isolated instance. Full tier detail is on the pricing page. A self-hosted VPS can be cheaper on paper ($4–6/month), but that figure hides the operational tax: you own the upgrades, the gateway config, and the 2am restarts.

Hermes or OpenClaw: which runtime should you pick?

Pick Hermes if you want a memory-first, self-improving single agent reachable over chat, Telegram, and Slack — the agent that grows with you. Pick OpenClaw if you need the broader surface: WhatsApp, cron, multi-agent routing, a workspace, or a Hosted Browser. You don't have to choose permanently — both run on one Lobsterland account and one bill, so evaluating is a launch, not a migration. If you've landed here from the launch news, the Hermes runtime support announcement covers what shipped, and the teams and browser-workflow comparison goes deeper on operating models.

FAQ

Is the Hermes agent runtime the same as the Hermes model?

No. The Hermes agent runtime is open-source software from Nous Research (MIT-licensed) that orchestrates memory, tools, channels, and model access into a persistent agent. The Hermes 2/3/4 family are separate fine-tuned language models. The runtime is model-agnostic: it can drive Claude, GPT, Gemini, or — via OpenRouter — Nous's own open-weight Hermes models. On Lobsterland you bring your own model key; the runtime is the agent, not the brain.

What can the Hermes runtime do on managed Lobsterland hosting?

On Lobsterland the Hermes runtime gives you dashboard and browser chat, Telegram and Slack channels (with per-user or whole-channel allowlists and an optional @mention requirement in groups), runtime logs to watch what the agent does, the agent's own self-written skills plus Hermes Hub installs, and bring-your-own model keys for Anthropic, OpenAI, OpenRouter, Google, and MiniMax. It runs as an isolated managed instance with no public IP and persistent memory across sessions, launching in about two minutes.

What does the Hermes runtime NOT do on Lobsterland?

On managed Lobsterland, Hermes does not include WhatsApp channels, a workspace or file editor, cron or scheduled work, multi-agent routing, a hosted browser, or usage analytics — those are OpenClaw surfaces. Hermes channels are Telegram and Slack only. If you need that fuller surface today, OpenClaw is the better fit; if you want a memory-first, self-improving single agent, Hermes is the one.

Can the Hermes runtime run Nous's Hermes models?

Yes — through OpenRouter you can point the Hermes runtime at Nous's open-weight Hermes models, which is the one place the runtime and the model family legitimately meet. It is not required: the runtime works equally well with Claude, GPT, or Gemini. On Lobsterland you choose the provider and bring your own key, and model usage is billed by your provider with no markup.