Voice input in the terminal: talk to your panes

The terminal finally accepts sentences. Rysh lets you speak a prompt into a pane — because agents take natural language, not just syntax.

For fifty years, the terminal has had exactly one input device: your keyboard. That made sense when everything you typed was syntax — flags, paths, pipes. Get one character wrong and the command fails.

Then agents moved in, and half of what we type into a terminal stopped being syntax. It's sentences. "Refactor this handler to use the new codec." "Why is this test flaky?" "Summarize what changed in the last ten commits."

Sentences are the one thing keyboards are slower at than your mouth.

So rysh supports voice input: you can speak a prompt into a pane instead of typing it.

Why voice belongs in a terminal now (and didn't before)

Voice-controlled shells have been demoed for decades, and they were always terrible for the same reason: shells want precision. Nobody wants to dictate tar -xzvf out loud. One misheard flag and you're extracting into the wrong directory.

Agents flipped that. In rysh, every pane is a conversation. Hit double-Esc and the pane toggles between shell mode (input goes to a real PTY) and prompt mode (input goes to the model — rysh runs on Claude, with your own API key). In prompt mode, the input is natural language by design. Natural language is exactly what speech is good at.

Voice input is a prompt-mode feature, not a shell-mode gimmick. You don't dictate flags. You dictate intent, and the agent turns intent into tool calls — every one of them visible in the pane, with the dangerous ones gated behind approval, same as if you'd typed the prompt.

What it looks like in practice

The flow is simple:

1. Focus a pane.
2. Make sure you're in prompt mode (double-Esc toggles it).
3. Trigger voice input and speak.
4. Your words land in the pane's input line as text — you can read them, fix them, then submit.

That last step matters. Voice input in rysh is dictation into a reviewable input line, not "speak and it executes." You see the transcription before anything is sent. The same trust model that governs everything else in rysh — see it before it happens — applies to your own voice.

Where it actually earns its keep

Hands-busy moments. You're reading a stack trace on one monitor, holding a coffee, mid-thought. "Explain why this panic happens when the pane group is empty" is faster to say than to type — and you don't lose your place in the trace.

Long, messy prompts. The prompts that produce the best agent work are often three or four sentences of context. People under-type context because typing is effort. Speaking three sentences costs almost nothing, and the agent's output quality follows the input quality.

Accessibility. For anyone with RSI or limited keyboard mobility, terminals have historically been one of the most hostile environments in computing. An agentic terminal that accepts speech is a genuinely different tool. This isn't a checkbox for us — panes are conversations, and conversations shouldn't require a keyboard.

Thinking out loud, literally. There's a real cognitive difference between composing a prompt and narrating a problem. "Okay, so the snapshot has the pane but the TUI renders stale output, which means either the KV write is time-gated or…" — dictate that stream into a pane, and you've got a rubber duck that answers back with a grep of your actual codebase.

What voice input is not

Honesty section, because we'd rather under-claim:

• It's not a voice-controlled shell. Shell mode is still keyboard territory, on purpose. Precision input deserves a precision device.
• It's not hands-free execution. Transcribed text sits in your input line until you submit it. Approval gates on dangerous tool calls still apply after that.
• It's not a replacement for typing. It's an additional input path for the input type — natural language — where speech beats keys.

And the standard rysh honesty footer applies: rysh runs on Claude with your own key, it's self-hostable on your own machines, and every claim in this post describes shipped behavior or is explicitly marked.

The bigger point

Voice input sounds like a small feature. It's actually a tell about what a terminal is becoming.

When the terminal was a syntax machine, keyboard-only made sense. Now that a pane can be a shell, an autonomous agent, a chat, or a live channel — the input layer has to match the conversation layer. Text when you want precision. Speech when you want speed of thought. The same pane, the same governance, either way.

Your panes can listen. Say something worth executing.

Rysh is early, and that's a feature: we're recruiting design partners — teams who want agents inside their real workflow, with a direct line to the founders. Apply at rysh.ai/design-partner.

More in this series: Double-Escape: one keystroke between your shell and your AI · Every pane is a conversation