Claude Code ships /voice. Codex has push-to-talk and realtime sessions. If you already have voice input in the agent you use every day, the obvious question is whether you need anything else.
Honest answer: sometimes not, and we would rather say so than pretend otherwise. We use Claude Code every day too, we use /voice, and we think it is good. Here is what the built-in features do well, what they leave untouched, and how to tell which case you are in.
- Native voice solves input. You speak, it becomes a prompt. That part is genuinely good.
- Nothing native answers you. You are still the one checking whether it finished.
- One agent, one project, no need for it to speak first? Native is probably enough.
What the built-in features actually do
They convert speech to a prompt, and they do it well. You hold a key, talk, release, and text lands where you would have typed it. Given that most people talk around three times faster than they type, that alone is worth using.
This is real voice coding in the narrow sense, and if nobody had built anything further it would still be an improvement over typing every prompt by hand.
It is also the half that is easy to build, which is why it arrived first and why it arrived in several places at once.
The half nothing native touches
Here is the thing the feature list will not tell you: faster input does not change what happens after you hit send.
The agent works. You wait. You have nothing to do for ninety seconds, so you tab away, and the ninety seconds becomes twelve minutes, because finishing made no sound. You come back to work that has been sitting there done.
Speaking your prompt instead of typing it saves you maybe thirty seconds. It does nothing at all about the twelve minutes. The bottleneck was never the sending.
Native voice input changed how the words get in. It did not change the waiting.
Three concrete differences
Beyond the reply, there are three things that only matter once you are past a single project:
- It works across agents, not inside one. Native voice belongs to the agent that shipped it. If you use Claude Code in one repo and Codex in another, you are learning two behaviours. A layer above them is one behaviour everywhere.
- Several projects at once. Each connected project can run a different agent and answer in a different voice, so a reply from a background repo announces itself instead of being ambiguous.
- Interrupting works. Full-duplex barge-in means you can talk over a spoken reply and still be heard. This only becomes relevant once something is speaking to you, which is exactly why native input does not need it.
Which case are you in?
Genuinely useful test, and it is short.
Native is probably enough if: you use one agent, work in one project at a time, sit at your desk while it runs, and are happy to check the terminal yourself.
You will want the other half if: you switch between agents, run more than one project, or keep losing time to the gap between an agent finishing and you noticing.
That second list is not a marketing construction. It is a description of what changes when the work stops being one conversation at a time.
Voice coding for Claude Code is free
Worth stating plainly, because people assume there is a catch: voice coding for Claude Code is free for life. Not a trial, not a seat count, not a word cap. Register once, no card.
You install it, type ski once inside your project, and Claude Code can hear you and answer you out loud. The speech recognition and the voice both run on your own machine, so there is nothing metered to meter. The only paid part of the product is sending an agent into a live video call, which is a separate thing entirely and not part of this.
We mention it because the honest comparison changes when one option costs nothing. If it were a subscription, “try both for a week” would be a real ask. It is not.
Use both, honestly
These are not mutually exclusive. Native voice input and a voice layer on top solve adjacent problems, and plenty of people will find the built-in feature covers everything they wanted. If that is you, genuinely, use it and ignore us.
What we would say is that the two halves feel different in a way that is hard to argue on a page. Speaking a prompt is a convenience. Being told the answer while you are stood at the window is a change in how the day is shaped, and most people do not believe that until it happens to them.
If you want to try the other half, it runs entirely on your machine, works with Claude Code, Codex, Cursor and around twenty other agents through one shared skill, and is free for life. Try both for a week and keep whichever you stop noticing.
More on how the loop is wired in how SKI works, or the argument behind it in voice coding in 2026.
