How to get Hermes Agent actually answering (2026)
Hermes Agent (Nous Research) runs on your machine. The usual failure is turning on the gateway before a plain chat replies. This flow follows the current quickstart: install → model → chat → doctor → extras.
The problem it solves
A mute Telegram bot is almost never «Telegram». It is a short-context model, a half setup, or a gateway stacked on a broken CLI.
What you will learn
- Install Hermes and confirm the binary exists
- Pick a provider and a model with ≥64K context
- Verify a chat that uses tools for more than one turn
- Diagnose with doctor before adding layers
- Leave gateway and cron for day two
Before you start
- macOS, Linux or Windows with a terminal
- A provider API key or willingness to use the Nous portal
- 18 minutes without installing Telegram at the same time
Steps 1
Install and reload the shell
On macOS/Windows, the Hermes Desktop installer is the short path. Via CLI: `curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash` then `source ~/.zshrc` (or bashrc). Confirm `hermes` exists (`which hermes`).
What you should see: The `hermes` command prints help or a version. Not «command not found».
Steps 2
Pick a provider before you ask for magic
Run `hermes setup` or, for zero API keys, `hermes setup --portal`. If you already have a key: `hermes model`. The model needs at least 64K context; otherwise the tool loop dies mid-sentence.
What you should see: A model selected and a note from you that the context window is ≥64K.
Steps 3
Verify a real chat, not the banner
Open `hermes` or `hermes --tui`. Ask something checkable: «Summarize this directory in 5 bullets and name the entrypoint.» Success = it replies, uses a tool if needed and lasts more than one turn. If you get empty output, stop: run `hermes model` again.
What you should see: A summary of the current directory, not an empty greeting.
Steps 4
Confirm sessions and diagnose
`hermes --continue` should resume the thread. If not, `hermes sessions list`. When it feels off: `hermes doctor` → `hermes model` → `hermes setup`. That order. Do not stack a gateway on top.
What you should see: doctor green or with a concrete failure you can read. The previous thread reopens with --continue.
Steps 5
Add the next layer only if the CLI already works
Then: `hermes gateway setup` for Telegram/Slack, `hermes skills browse`, `hermes tools` to tighten permissions. To keep raw commands off your Mac: `hermes config set terminal.backend docker`.
What you should see: Local chat still works after the extra. If local breaks, undo the extra.
Real use cases
Summarize a local repo
The agent lists the tree, names the entrypoint and stays in that directory.
A maintenance script
One job, one directory, tight permissions. Gateway later if needed.
The same job as Cursor, outside the IDE
When the change is no longer a diff but an agent that lives in the shell.
Common mistakes
Gateway on day one
If the CLI does not talk, Telegram will not either. Doctor first.
An 8K or 32K model
The tool loop falls over. Raise context before you blame Hermes.
Open permissions «so it works»
Tighten tools after the chat already answers, not the other way around.
Close
Hermes is an agent on your machine, not an IDE and not a cloud teammate. One provider, one tool-using chat, doctor green. Everything else can wait.
Next steps
- Run a second local job tomorrow.
- If the work is a repo diff, go back to Cursor.
- If the work is a cloud browser, that is Grok Bot.
Free guide
Want the 15-tools guide?
We send it free. We will also ping you when the next guide in this path goes out.
Unsubscribe whenever you want. We do not sell the list.
Part of the path Build with agents