Heralding Herald

Giving your coding assistant a voice

I’ve been using Claude Code as my primary development tool for a while now. Voice in, terminal out. I talk, it works. But there was always a gap in the other direction. I could see what it was doing if I watched the screen, but I didn’t always want to watch the screen. Sometimes I’m leaning back, thinking. Sometimes I’m looking at something else. Sometimes I just want to know it’s still moving without staring at a scrolling terminal.

So I built Herald. It’s a real-time spoken narrator for Claude Code. As Claude works, Herald announces what it’s doing in short, natural sentences, spoken aloud through your speakers. “Reading the config file.” “Searching for test failures.” “Updating the retry logic.” Like a play-by-play announcer for your AI pair programmer.

The Bigger Picture

If you buy the premise that voice is replacing the keyboard, then you have to follow the thread. The input side is already there. I talk to my computer and things happen. But the output side is still stuck in the old world. I speak, and then I read. That’s half a revolution.

Think about where this goes. The keyboard disappears. The mouse disappears. Eventually the screen disappears too, or at least stops being the primary channel. What replaces it? Audio. And after audio, maybe something more direct. But audio is the next step, the one we can build right now.

Herald is a small, simple proof of that idea. It closes the loop. Voice in, voice out. I tell Claude what to do, and it tells me what it’s doing. No screen required. I can lean back, close my eyes, and still know exactly where things stand. “Checking if the tests pass.” “Updating the config.” “All done. The migration ran successfully.” That’s not a gimmick. That’s what the interface wants to become.

Why Audio Changes the Experience

When another person works next to you, you get ambient awareness for free. You hear them typing. You hear them mutter “what the hell” when something breaks. You don’t have to ask “what are you doing?” because the environment gives you enough signal.

Claude Code doesn’t give you that. It works in silence. You can scroll up and read the tool calls, but that’s active, directed attention. Hearing a voice requires almost none. It’s the same information, delivered in a way that costs less attention.

When I can hear what Claude is doing, I let it run longer without checking. I trust it more. Not because the narration adds information I couldn’t get from the terminal, but because spoken language is easier to process passively. Reading scrolling text requires focus. Listening to a calm voice saying “Searching for leftover TODOs in the codebase” does not.

And there’s something else. This matters for people who can’t see the screen at all. If you have a vision impairment, a terminal full of scrolling tool calls is not an interface. But a voice that narrates what’s happening? That’s an interface. Herald wasn’t built as an accessibility tool, but the moment you make the output audible, you’ve opened the door to anyone who works better with their ears than their eyes. That feels important. AI-assisted development shouldn’t require staring at a screen to be useful.

How It Works

Herald hooks into Claude Code’s event system and converts each action into a short spoken sentence via text-to-speech. When Claude finishes a turn, it summarizes what was accomplished. When Claude needs your attention, it plays a ding. The whole thing is three files and takes about two minutes to set up.

The repo is at github.com/rodionsteshenko/herald. Do whatever you want with it.

The Direction

The interesting thing about Herald isn’t the implementation. It’s the direction it points toward.

Right now, AI coding tools are screen-first. You watch text scroll, you read diffs, you click approve. That works. But it’s a transitional state. The natural endpoint of voice-driven development is voice-driven feedback. You shouldn’t need to look at a screen to know what your assistant is doing any more than you need to look at a screen to know what your coworker is doing.

Herald is a rough sketch of that future. Someone will build a much better version. Someone will build it into the tools themselves. I’m just here to say: the output side of the interface is lagging behind the input side, and closing that gap changes the experience more than you’d expect.

Voice in, voice out. That’s the whole idea.