Hermes Agent end to end: a VPS install, Hermes Desktop on a remote gateway, memory and SOUL.md, skills, and Bot Mode with group chats

What the video shows

• Why a server. He runs Hermes on a virtual private server rather than his own computer: it stays online, has a firewall and can be wiped without risking hardware. He deploys a Hostinger template (the KVM 2 plan) that sets Hermes up in a Docker container and shows the admin username and password in the container's environment tab.
• Three ways to manage it. The browser dashboard, a terminal inside the container (run hermes for the text UI), and a coding agent installed on the same server, so that if Hermes breaks you can ask it to fix the install.
• Model. He runs hermes setup model, signs in with a ChatGPT subscription through a browser code, and picks a GPT-6.1 Sol model at medium reasoning.
• Desktop app. He installs the desktop app, opens the settings menu, goes to Gateways, adds a remote gateway with the server's URL, signs in with the same username and password, and sees the earlier web chats appear. He warns that using Desktop without connecting gives you a separate local agent with access to your own files.
• Memory and personality. A global USER.md holds facts about you and a MEMORY.md holds Hermes's own long-term notes, both loaded at the start of a chat. SOUL.md sets personality, like a system prompt; Desktop's Manage profile view shows it. He asks Hermes to rewrite it (name, directness, "never contact another person without asking me first") and the name changes.
• Skills. Hermes saves a skill after a task and updates it from feedback; /learn asks it to learn from what it just did. A Capabilities page lists skills and tools, and he warns to read hub skills before installing because some may be malicious. He also tours connectors, the plugin catalog and messaging (a Telegram QR code, Discord and Slack tokens).
• Bot Mode. New Agent in the Bots tab creates a bot with its own name, model and SOUL.md ("life coach" on a cheaper model). He adds two bots to a group chat, writes @all, and they reply to each other. He also asks for a recurring greeting and Hermes creates a scheduled job.

Checked against the Hermes docs

From the model, multi-connection Desktop and Bot Mode pages, read 2026-10-27.

• The model command. The docs do not list hermes setup model. hermes model is "the canonical way to switch defaults" and writes model.provider and model.default to ~/.hermes/config.yaml; saving a model then offers a reasoning-effort step (minimal through ultra, written to agent.reasoning_effort). Inside a chat, /model <name> [--provider X] [--global | --once] switches for the session, persistently or for one turn. The page does not describe a ChatGPT or Codex sign-in flow beyond saying hermes model includes browser OAuth flows.
• Remote gateway sign-in. The docs path is Settings, Gateways (Cmd or Ctrl+comma), Add connection, Remote gateway, with a unique name and the base URL of a running hermes serve backend. Authentication is a session token or Nous Portal OAuth. Username and password are not mentioned, so the login in the video comes from the Hostinger template, not from Hermes itself. Click Test: it checks both HTTP and WebSocket, and a WebSocket failure usually means a proxy, firewall or origin guard blocking /api/ws. The backend must already be running; Desktop does not start it for a remote gateway.
• Bots. Open the Bots tab and click New Agent; the quick path asks for Name, Title and Description, and Advanced lets you clone a profile, pin a model and provider, write a SOUL.md and pick skills, toolsets and MCP servers. Each bot is a Hermes profile under ~/.hermes/profiles/<name>/ with its own config, memory, skills, credentials and chat history. API keys are copied from the main profile by default, but single-use OAuth logins are not.
• Group chats and their limits. A group takes 2 to 6 bots. Each message you send triggers up to three serial rounds of replies, with hard caps of 10 messages per send and 3 rounds. @all or @everyone releases held members, stop words can hold a member, and bots escalate to you with @user. The back-and-forth in the video therefore ends by itself.
• Not located in the docs. We did not find /learn documented in the docs index, and we did not verify the USER.md and MEMORY.md behaviour on the Memory page; treat those as the video's description.

Gotchas & Caveats

• Putting a coding agent on the same server means it can run commands there. Sign it in on a server you are happy for it to control.
• A remote gateway token or login gives full control of the agent. Keep the URL private and prefer the documented token or OAuth sign-in.
• Check hub skills before installing, as he says. A skill is instructions plus scripts your agent will follow.
• He is candid that he has not explored how group chats end; the documented caps are the safeguard.

Key Takeaways

• Use hermes model to set the provider and reasoning effort, not hermes setup model.
• Connect Desktop to your server with Settings, Gateways, Remote gateway, and press Test before chatting.
• Give each bot its own profile, model and SOUL.md, and rely on the group-chat caps (10 messages, 3 rounds) rather than hoping they stop.