The short answer
If you're walking, commuting or cooking and want to send your coding agent one more instruction, the practical approach is to let your phone turn speech into text and deliver that text to the session already running on your computer. Your phone keyboard's dictation key works in any text field, including an SSH client, but a long transcript squeezed into a terminal prompt is hard to check and awkward to fix before you hit enter. BotBus puts voice input directly into the boxes you use to follow up, answer an agent's question and start a new task: on iPhone and Android you tap the mic, watch the words appear, edit and send; on Apple Watch you raise your wrist, tap, speak, and the follow-up goes out.
Voice fixes the "typing is hard" problem. It doesn't fix the "I can't see enough" problem: reviewing a diff or approving a risky command still needs your eyes. The rest of this post covers both.
What already works, and who it suits
Keyboard dictation on your phone. Both iOS and Android keyboards have a mic key that works everywhere, SSH apps included. Nothing to install. The friction is the terminal: the transcript lands on a single command line, fixing a misheard filename means dragging the cursor around, and the connection often drops when the app goes to the background. Good enough if you already keep an SSH session open and only need the occasional sentence.
Desktop dictation tools. When you're at your desk and just tired of typing, OS dictation and the many desktop voice-typing apps work well straight into an agent's terminal or desktop app. They don't help once you've left the desk.
Chat bots and messaging gateways. Some people wire their agent into a chat app and send voice notes that get transcribed and forwarded. It works, but there's an extra hop, and your audio and conversation pass through that platform. We compared the trade-offs in chat bots vs a mobile app for coding agents.
BotBus voice input. Useful if you already want to keep an eye on your agents from your phone or wrist and just need to reply or kick off a task without opening a terminal. It's not a voice assistant; it's dictation built into BotBus's own input fields.
Where you can speak in BotBus
Requirements first: BotBus runs on the computer where your agents live (a menu bar app on macOS 26 or later, plus Linux and Windows versions), and your phone is paired with it. The computer must be on and online. Whether you can follow up, answer or start a task depends on the agent's connector and the task's current state; when an action isn't available the button is disabled and the app tells you why.
iPhone and Android:
- Follow-ups. The input bar at the bottom of a task has a mic button. Tap once to start, and the words stream into the field as you talk; tap again to stop.
- Answering questions. When the agent stops to ask you something, the same field is your reply box. Multiple-choice questions are just taps.
- New tasks. The new-task screen has the mic too. Pick the computer, project and agent, then say what you want done.
Nothing is sent automatically. The transcript sits in the field so you can type more, fix a misrecognized word, then press send yourself; what goes out is exactly what you see. If there's already text in the field, new speech is appended: Chinese and Japanese join directly, while English words get a space so two phrases don't fuse into one word.
Apple Watch:
Speaking is the easiest input on a watch, so the main button opens system dictation in one tap. Next to it is a keyboard button for Scribble or typing when you can't talk out loud.
- Follow-ups and replies. Tap "Follow Up" in a task (it becomes "Reply" when the agent is waiting on an answer), speak, confirm in the system dictation sheet, and it's sent. No extra step.
- New tasks. Dictation fills the task description. You can speak in several passes, each appended to the last, and swipe the row to clear it.
- Approvals and multiple-choice questions. These are buttons, not voice. Approve and Deny are pinned to the bottom so you don't have to scroll.
For the full watch experience, see managing AI coding agents from Apple Watch.
Where recognition happens and what permissions it needs
Prompts to a coding agent often contain code, file names and paths, so it's worth knowing where your voice goes.
- iPhone uses Apple's speech recognition framework. When the device supports on-device recognition, BotBus requires it; otherwise the system sends audio to Apple's servers, which cap a single request at roughly a minute, so split long requests. The first use asks for microphone and speech recognition permission. The recognition language is the first one in your system language list that the recognizer supports.
- Android uses the system speech recognizer, preferring on-device and offline recognition when available. On Android 13 and later BotBus checks which languages the recognizer supports and matches your system languages; on older versions it leaves the choice to the recognizer.
- Apple Watch uses watchOS's built-in dictation. BotBus doesn't record audio itself and doesn't need its own mic permission on the watch.
Once it's text, it travels exactly like typed text: encrypted on your phone, relayed by BotBus, which only sees ciphertext, and decrypted on your computer. See the security overview for the precise boundary: content is end-to-end encrypted, but the server can see device IDs, online status, update times and data sizes. One more distinction: once the agent on your computer receives your prompt, it sends it, along with context, to its model provider, and that leg is outside BotBus's encryption. That's why each device shows a disclosure the first time you send something to an agent through BotBus, asking you to check the provider and its privacy policy before continuing.
What kind of prompt works well spoken
Dictation shines for short, unambiguous instructions. A few habits that help:
- State the goal and the constraint, not the edit. "Translate the login error messages into Spanish, don't touch the API" is easier to say and to recognize than a line-by-line description. Let the agent find the code.
- Be careful with identifiers. Function names, paths and flags are where recognition fails most. On the phone, glance at the field and fix them before sending. On the watch, follow-ups go out right after you confirm, so prefer descriptions like "the config file you edited last time."
- Split long requests. Rambling for three minutes invites dropped words, and server-side recognition on iPhone hits its time limit. Say the goal, check it, then add constraints.
- Lead with the decision. If the agent asks "option A or B?", start with "B", then explain. A garbled reason won't flip your answer.
- Use it to start things. An idea on a walk is a perfect new task: describe it, let the agent read the code and propose a plan, and review properly when you're back. More in starting a coding task from your phone.
Leave long code snippets, regexes and anything that depends on exact quotes or indentation for the keyboard, or let the agent write them.
Limits
- Voice is only an input method. It can't do anything typing can't. Follow-ups, answers and new tasks depend on the connector, the task state and whether the computer is online; a read-only observed task won't accept your words either.
- No voice approvals. Approving or denying is a tap, and so are multiple-choice answers. That's deliberate: you should see the command or file change before you approve it.
- Accuracy depends on the OS. BotBus uses the built-in recognizers of iOS, Android and watchOS, not its own model. Accents, mixed languages and background noise all matter, and devices without on-device recognition need a network connection.
- Watch follow-ups send on confirm. There's no character-level editing on the watch. If something came out wrong, send a quick correction right after.
- No spoken replies. BotBus doesn't read the agent's answers aloud; you still read them on screen.
- Platforms. iPhone and Apple Watch apps are on the App Store (iOS 26 / watchOS 26 or later). The Android app needs Android 8.0 or later and is available from the download section.
FAQ
Does voice input cost extra or need another app?
No. It's part of the BotBus phone and watch apps, and recognition uses the system's built-in service.
Can BotBus's servers hear what I say?
BotBus doesn't upload recordings. iPhone and Android prefer on-device recognition; when an iPhone can't recognize on-device, the system sends audio to Apple. The resulting text is encrypted on your phone before it passes through the BotBus relay. On your computer, the agent forwards it to its model provider under that provider's privacy policy.
Can I dictate in languages other than English?
Yes. Recognition follows your phone's system languages, and the agent receives whatever text was transcribed. Claude Code and Codex handle instructions in many languages; just double-check English identifiers inside non-English speech.
Can I approve a command by voice?
No. Approvals and multiple-choice questions are buttons on both phone and watch, so you see what you're approving. If you approve the same kind of action all day, read how to use auto-approve safely.
Why can't I edit a watch follow-up before it's sent?
A watch screen is too small for word-level editing, so a follow-up is sent once you confirm it in the dictation sheet. Re-dictate before confirming, or use the keyboard button to Scribble instead. New tasks are different: you can dictate in several passes and clear the text before starting.