---

Step 1: Pick the right machine

Hermes can read, change and delete files, and it can write its own skills. So do not install it on the laptop with your private data. Use a machine with nothing on it that you cannot afford to lose: a small server, a Mac mini on a shelf or a virtual machine. Mine runs on a Mac mini.

macOS, Linux, Windows (native or via WSL2) and Android via Termux are supported (installation).

Step 2: Install Hermes

On macOS, Linux and WSL2:

curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash

On Windows, in PowerShell:

iex (irm https://hermes-agent.nousresearch.com/install.ps1)

Then reload your shell (source ~/.zshrc or source ~/.bashrc). If you prefer clicking, the website has a desktop installer for macOS and Windows.

Step 3: Choose a model

Hermes is the frame, you pick the model. Start the setup with:

hermes setup

You can choose Quick Setup through Nous Portal (one login, billed by Nous), Full Setup with your own keys, or Blank Slate with extras turned off. I start with Quick or Full. Blank Slate switches off memory, among other things, and you do not want that here.

Switch the model any time with:

hermes model

The list includes Nous Portal, OpenAI Codex, Anthropic and OpenRouter, among others. Important: the model needs at least 64,000 tokens of context, smaller ones are rejected at startup. For local models set it with --ctx-size 65536 (llama.cpp) or -c 65536 (Ollama) (quickstart).

I linked Hermes to my ChatGPT subscription. A cheap model is enough for learning.

Step 4: The first chat

hermes --tui

Give it something you can check, for example: "Summarize this repo in five bullets and tell me where the entry point is." If it answers without errors and carries on after the first turn, the base works. Only then add Telegram, jobs and skills. hermes --continue resumes the last session, which tests the memory.

Step 5: Connect Telegram

So you can message Hermes on the go:

1. In Telegram, open @BotFather and send /newbot. Pick a name and a username (it must end in bot). You get a token. Treat it like a password.
2. Get your numeric user ID, for example from @userinfobot.
3. Run hermes gateway setup and choose Telegram. Alternatively put TELEGRAM_BOT_TOKEN and TELEGRAM_ALLOWED_USERS into ~/.hermes/.env.
4. Start the gateway with hermes gateway and send the bot a message.
5. Send /sethome in the chat where you want your job results.

Always set TELEGRAM_ALLOWED_USERS. Without that list the gateway rejects everyone, and that is good, because your bot has access to a machine. To keep the gateway running after a restart, use hermes gateway install (Telegram setup, gateway).

One caveat the docs do not mention: Telegram bot chats are not end-to-end encrypted. Do not send the bot passwords or data you do not want stored at Telegram.

Step 6: Create your first job

You describe it in plain words, for example straight in the chat: "Every morning at 9 find the most important AI news and send me a summary on Telegram." On the command line:

hermes cron create "every 1d" "Summarize the most important AI news" --paused --paused-reason "check first"

Create the job paused, run it once with hermes cron run <name>, read the result and then switch it on with hermes cron resume <name>. Results go to your home chat with --deliver telegram. All time formats and targets are in the cron docs.

Step 7: Watch the costs

Hermes uses a lot of tokens because it drags context, memory and tools along. Compared with an assistant you hire, it is still a fraction. That holds as long as you use one model. One creator burned about 130 dollars in a single day by running several large models in parallel (Metics Media, YouTube). My advice: start with one model, watch the costs for a week and only then build multi-model setups.

Step 8: Connectors, honestly

Here Hermes is not as far along as ChatGPT or Claude. Many connectors are more mature there. My experience: when Hermes runs with a model whose connectors are already hooked up, it works very well. The Google connection, on the other hand, costs me nerves.

What the docs list as integrations:

• MCP servers for external tools, over stdio or SSE (docs)
• Web search via Firecrawl (default), SearXNG, Brave, DuckDuckGo, Exa and more
• Browser automation via Browser Use, Browserbase or local Chromium
• Messengers: Telegram, Discord, Slack, WhatsApp, Signal and about 20 more

Which of them are best for you depends on your day. Start with one: a messenger and web search are enough to begin.

Step 9: Move over from OpenClaw

If you already use OpenClaw, Hermes has its own migration command:

hermes claw migrate --dry-run

The dry run only shows a preview and changes nothing. The tool reads ~/.openclaw, imports memory, SOUL.md and skills, among other things, and does not modify the OpenClaw files. Make a backup anyway, and check skills with complex logic by hand after the import. The exact flags differ by version: hermes claw migrate --help shows what your install supports (migration skill).

Common mistakes

• Installing Hermes on your main machine. It can change and delete files.
• Not setting TELEGRAM_ALLOWED_USERS. Then the bot answers nobody, or you open it to everyone.
• Choosing a model under 64K context. The start aborts.
• Creating jobs active right away. Test paused first, then switch on.
• Several large models in parallel. That is where the bills come from.

Next Steps

1. Install Hermes on a machine without private data.
2. Get a normal chat working before you add anything.
3. Connect Telegram and create a paused job.
4. After a week, look at costs and results.