Using Hermes: CLI, TUI, Slash Commands, and Sessions

Using Hermes From the Terminal#

Hermes is a terminal-native tool. There's a classic readline CLI and a richer TUI, but they share the same commands, sessions and data. This lesson is the practical 'what do I type, what happens next' tour — launching, interrupting, resuming, running things in the background, and keeping context under control.

Primary source: the official CLI user guide at https://hermes-agent.nousresearch.com/docs/user-guide/cli. This lesson is a summary in our own words; refer to the upstream doc for the authoritative command list.

Two Interfaces, One Session#

The classic prompt (hermes) is a simple REPL that works over SSH and in minimal environments. The TUI (hermes --tui) adds modal overlays, mouse support, a persistent status bar, and a session-resume panel that shows a compact recap of the previous conversation when you rejoin.

Critically, both write to the same SQLite store at ~/.hermes/state.db — sessions created in one interface are visible from the other. Switch freely.

The Status Bar and Adaptive Layout#

The persistent status bar above the input area tells you, at a glance, things you'd otherwise have to ask for:

• Active model and provider.
• Token usage and context-fill percentage — watch this before long operations.
• Estimated running cost for the session.
• Elapsed session duration.

The layout adapts to terminal width — the full bar appears at 76+ columns and degrades gracefully for narrower windows. If you spot the bar collapsing unexpectedly, resizing the window back usually restores it.

Launching and One-Shot Queries#

The useful invocations to remember:

hermes                               # interactive session
hermes chat -q "what's my disk usage?" # one-shot, non-interactive
hermes --tui                          # modern TUI
hermes --continue                     # resume most recent session
hermes --resume <session_id>          # jump to a specific session
hermes -s my-skill                    # preload a skill at launch
hermes --model anthropic/claude-sonnet-4 --provider nous

The -q (one-shot) form is particularly useful from other scripts — Hermes prints the model response to stdout and exits without opening the REPL.

Slash Commands#

Inside a session, anything starting with / is a slash command, not a message to the model. Press / and autocomplete lists them. The ones you'll use constantly:

• /help — the canonical list, with descriptions.
• /model — switch model mid-session (the session keeps its context).
• /tools — see which toolsets are currently enabled.
• /skills browse — explore the official skills hub.
• /title — name the session so it's findable later.
• /voice on and Ctrl+B — start voice input and trigger a recording.
• /reasoning high — turn up model reasoning effort for hard problems.
• /compress — summarise older turns when the context is filling up.
• /usage — detailed token + cost breakdown.
• /verbose — cycle display verbosity (off → new → all → verbose).

Multi-Line Input and Pasting Code#

Accidentally sending a half-written prompt is the most common early frustration. Hermes gives you three tools:

• Press Alt+Enter or Ctrl+J to add a newline instead of submitting.
• End a line with \ to continue onto the next line.
• Pasted code blocks are auto-detected and buffered as a single message, not sprayed into multiple sends.

Interrupt, Don't Start Over#

When the agent goes off-track, the reflex should be 'interrupt and redirect', not 'quit and restart':

• Type a new message while it's working — that interrupts the current operation and takes the new instruction into account.
• Single Ctrl+C — interrupts the current turn, keeps the session alive.
• Double Ctrl+C within 2 seconds — forces an exit.
• Ctrl+Z — suspends to background on Unix (fg to resume).

Because interruptions are tracked, the next prompt you send has the context 'you were doing X, I stopped you, now do Y instead' — the model knows why it was cut off.

Background Tasks#

/background is one of the most underrated features. It spawns a completely separate agent session as a daemon thread, sharing your configuration (model, provider, reasoning) but keeping its conversation isolated from your main thread. Each background task gets an ID — check on them later without cluttering the foreground conversation.

Typical use: 'investigate this bug in the background while I keep working on the feature'. The main session stays focused, the background thread is there when you want the results.

Context Compression and Quick Commands#

Two ergonomic features worth knowing early:

• Context compression happens automatically as the window fills — Hermes preserves initial and final turns and summarises the middle, so long sessions don't just fall off a cliff.
• Quick commands, defined in ~/.hermes/config.yaml under quick_commands, map short names to shell operations that run without invoking the LLM — useful for things you don't want to pay tokens for.

Next Steps#

You now know enough CLI to work comfortably. Lesson 3 covers the four features that make Hermes interesting as a platform, not just a chat tool — memory, skills, toolsets and voice. If you prefer to run Hermes from your phone, skip ahead to Lesson 4 on messaging platforms.