The questions people actually ask before and after installing Hermes. Honest answers — not the softball version from the official docs.
Three components:
Total for moderate use: roughly $15–$25/month. Comparable to a single ChatGPT Plus subscription, but with always-on availability and self-hosting.
Set a hard spending cap on your OpenRouter key when you create it. This is the most important step. If your key hits the cap, it stops working — no further charges. You can set a monthly reset cadence so the cap automatically restores on the 1st of each month.
Also: don't run your most expensive scheduled jobs on the highest-tier model. Use Opus or GPT-5 class only for one-off heavy tasks; use a Sonnet or mini-class model for daily briefings and routine requests.
OpenRouter gives you a small amount of free credits when you sign up, enough to test several conversations. You'd still need a VPS for a real deployment, but you can test the setup with minimal spend on day one. Some VPS providers also offer trial credits for new accounts.
Minimum: 2 vCPU, 8 GB RAM, 100 GB SSD with Docker preinstalled (or use a one-click Hermes Docker template).
Most starter KVM plans at Hostinger, DigitalOcean, Linode, Vultr, or Hetzner meet this. If you're new to VPSes, the one-click Hermes template on Hostinger is the lowest-friction option — it handles Docker setup automatically.
Yes, but it defeats the purpose. Hermes works on your laptop for testing — but when your computer sleeps or restarts, the gateway goes offline and scheduled jobs stop running. The always-on value comes from the VPS. For production use, a VPS is required.
Not deeply. The setup involves: opening a web terminal, navigating to a directory (cd), running two commands to enter the container, and running hermes setup which is a guided wizard. If you can copy and paste a command into a terminal, you can do it.
Ongoing maintenance is minimal — mostly occasional gateway restarts after VPS reboots and model updates. None of it requires Linux expertise.
Anthropic (Claude models), OpenAI (GPT models), and OpenRouter. OpenRouter is the most flexible: one key, 200+ models from every major provider, and you can switch between them with a single chat command (/model). It's what most people use.
Via the Gateway, Hermes connects to: Telegram, Discord, Slack, WhatsApp, Signal, and 10+ additional platforms. Telegram is the easiest to set up — no developer portal, just two quick conversations in the Telegram app (BotFather for the bot token, UserInfoBot for your numeric ID).
Yes. You can connect as many platforms as you want via hermes gateway setup. It's the same agent on every platform — same memory, same skills, same history. Switching between Telegram on your phone and Slack on your laptop doesn't create separate agents; it's one agent accessible from multiple channels.
Nothing. During setup, you give Hermes an "allowed user IDs" list — only those Telegram numeric IDs can interact with the bot. Anyone else who finds the bot username and messages it gets silently ignored. The bot appears to not exist to anyone not on your list.
Four security features are on by default:
The main risk surface is your API key. If someone gets your OpenRouter key, they can drain your credits. Keep it private — never paste it in chat, tickets, or screenshots. Set a spending cap so even a leaked key has a hard ceiling.
Your conversation history, skills, and memory store all live on your VPS — not in a Hermes cloud. The VPS provider has access to the machine (standard hosting terms), but Hermes the company doesn't receive your conversations.
Your messages do pass through your AI provider (OpenRouter, Anthropic, or OpenAI) as API calls. Check your chosen provider's data retention policy if that matters for your use case. OpenRouter has a setting to disable conversation logging.
Not in week one. The recommended approach: start on Telegram only, pick one or two workflows, and run them for a few weeks to build trust with how Hermes handles approvals and decisions. Once you know how it behaves, you can extend access to more sensitive systems with confidence. Rushing to connect email, banking, or a password manager on day one is how you end up with an agent making decisions you didn't expect.
Built-in memory is a structured store on your VPS. When you tell Hermes something (your name, a preference, a fact about your business), it saves it as a retrievable entry. On future conversations, it loads relevant memories as context before responding.
You can inspect the full memory at any time: "Show me what you remember about me." You can add to it, correct it, or delete specific entries by asking directly.
Yes. You can delete individual memory entries by asking ("Forget that I mentioned X"), or you can wipe the entire memory store from inside the container. The memory files are on your VPS, so you can also back them up before wiping if you want to keep a snapshot.
Technically yes — you can add multiple Telegram user IDs to the allowed list. In practice, shared use gets complicated quickly: one agent, one shared memory store, one set of preferences. Hermes is built for personal-assistant use by one person. If you need team access, it's better to spin up separate agents on the same VPS (one container per user) than to share one.
Four structural differences:
The trade-off: Hermes requires a VPS and some setup. ChatGPT is instant and has no maintenance overhead.
Both are self-hosted AI tools, but they solve different problems. OpenClaw is a gateway that connects your existing apps (chat tools, files, APIs) to an AI agent — the emphasis is on integrations and workflow automation across your business stack. Hermes is a personal assistant you talk to directly via Telegram or Discord — the emphasis is on conversation, memory, and scheduled tasks for one person. They can complement each other rather than compete.
They serve different use cases. n8n is a visual workflow automation platform — it's excellent for structured pipelines that connect apps on a trigger (form submitted → send email → update CRM). Hermes is a conversational AI assistant that learns your preferences and responds to natural-language requests. If you have structured workflows with clear triggers and outputs, n8n is the better tool. If you want an AI you can talk to that also runs scheduled tasks, Hermes fills a different gap.
The gateway needs to be running. Re-enter the container and run:
cd /docker/hermes-agent-XXXX
docker compose exec -it hermes-agent /bin/bash
hermes gateway runTo prevent this from happening again, make sure the gateway is started as a background service rather than an interactive process. Some installs require this manual step on first boot; after that it should auto-start on VPS restarts.
First, check that your home channel is set correctly. Ask Hermes: "What is my home channel set to?" If it's wrong, re-run hermes setup and confirm your user ID as the home channel. Second, verify the job ran: "Show me the last run time and output for job [name]." If the job errored, it'll tell you why.
Check two things: (1) What model is set as your default? Premium models (Opus, GPT-5 class) cost 10–50x more per message than mid-range models. Switch your default to a Sonnet- or mini-class model: send /model and change it. (2) Are any scheduled jobs using a premium model? Ask Hermes to list all jobs and their model settings.
Yes. Your skills and memory are files on the VPS. Before migrating: take a snapshot of the current VPS (your host's dashboard), or manually export the skills and memory directories. Redeploy on the new server, then copy those directories back. All your customizations and history will be intact. The setup wizard only needs to run once; subsequent moves just need the data files.
Was this guide helpful?