Argoforgebeta
Argoforgebeta
Hermes reads the start of each Claude Code answer aloud. You can say your next message, and it goes straight into the terminal. The Argoforge app does it all on your Mac; the Argoforge server takes no part.
Step by step
Two parts: answers read aloud and dictation. Both run on your Mac, unless you pick an online voice or your Mac can't recognise speech on its own.
The app picks up the answer. It stays silent when voice is off and in scripted runs (claude -p or the SDK).
Code, tables, links and file paths are dropped, with the punctuation they leave behind. At most two sentences remain, up to about 200 characters; if the answer ends with a question to you, that question is read too.
OpenAI (with your own key) or Microsoft's free voice, edge-tts, turns the text into sound. With the built-in macOS voice this step is skipped.
A new answer from the same session cuts off the one still playing; one from another session waits for it to end (20 seconds at most) and starts with its project's name when the last answer spoken was another project's. If the online voice fails, the built-in macOS voice reads the text instead.
Hold the shortcut (⌃⌥Space by default) and speak, or press your AirPods if you set that up. An answer being read aloud stops at once, so the microphone does not hear it. macOS turns speech into text, on the Mac itself when it can. The app pastes the text into the last terminal you used; in a Claude Code session it also presses Enter, elsewhere the text waits for your Enter.
Your data
Hermes uses neither Claude nor any other language model. Only if you pick the OpenAI voice does OpenAI's speech model read the text, with your own key. The Argoforge server runs nothing for Hermes.
Rules
Two sentences, about 200 characters, plus a closing question if there is one. Code and file paths are never read aloud.
An answer from another session never cuts one off: it waits its turn and names its project when the project changes.
When you start dictating, an answer being read aloud stops at once, so the microphone does not hear it.
Runs started with claude -p or the SDK are never read aloud.
Voice and dictation start switched off. Answers are read aloud only while Argoforge itself is on.
If the online voice fails or takes longer than 20 seconds, the built-in macOS voice speaks instead — a Russian voice for Russian text, the system voice otherwise.
Microphone, Speech Recognition and Accessibility (to paste into the terminal).
The text is pasted via the clipboard. Enter is pressed for you only in a Claude Code session, so a misheard phrase never runs as a command in a plain shell. Whatever you had copied — text, a picture or a file — then comes back.
How to use
The Listen button on the Hermes page plays a sample, even before you switch voice on.
The app installs it with one button. It needs Python 3 and the internet.
Dictated text goes to the last terminal app you used: Terminal, iTerm2, Ghostty, Warp, kitty, WezTerm, Hyper or Alacritty.
Dictation can use a language other than the app's: set it on the Hermes page. Developer terms (git, commit, Swift…) and the names of your projects are passed to the recogniser as hints; project names only when your Mac recognises speech on its own, so they never leave it.
The Hermes page shows which of the three are missing.
A short retelling of the answer instead of its first sentences is planned. Until then answers are always shortened by fixed rules.