Blog22 September 20263 min read

Every AI lab now does voice. None of them talk to each other.

Claude Code got a /voice command. Cursor put a mic next to its Agent box. ChatGPT learned to grab your screen with Appshots. The labs agree that talking and showing beats typing. They just each built it for their own app, and only their own app.

localhost:3000/settings1this button, under the fieldany agent

What actually shipped, tool by tool

Claude Code: /voice

Anthropic shipped voice dictation for Claude Code in March 2026. You run /voice, hold Space, talk, and let go; your words land in the prompt as editable text. It is tuned for developer words like OAuth and localhost, and it does not eat your usage limits. (A good walkthrough of /voice.)

What it is not: it does not see your screen. It turns speech into text, inside Claude Code, and that is the whole job.

Cursor: the mic in the Agent box

Cursor added a microphone next to the Agent input, with Space as the shortcut on macOS. Handy, with one catch people keep tripping on: it dictates into the Agent prompt, not into Chat, inline edit or the terminal. (Details on its limits.)

ChatGPT: Appshots

OpenAI's desktop app can now grab the window you are looking at, pixels plus whatever text the app exposes, and drop it into a chat. Press both keys on a Mac and the context comes along. (OpenAI's Appshots docs.)

It is the closest thing to "show, don't type" from a big lab. It also only ever goes to ChatGPT.

The pattern: everyone is building the same thing, in a walled garden

Look at the three together and the direction is obvious. Every lab has figured out that the bottleneck is not the model, it is you describing the problem. So they are all racing to let you talk and show instead of type.

But each version is welded to its own product:

Hears youSees your screenKnows which elementWorks in other tools
Claude Code /voiceYesNoNoNo, Claude Code only
Cursor voiceYesNoNoNo, Cursor's Agent box only
ChatGPT AppshotsVia chatThe whole windowNoNo, ChatGPT only
DeikoYesWhat you point at, croppedYes, with its labelYes, any agent or chat

If you use one assistant for everything, forever, that is fine. Almost nobody does. You bounce between Claude Code in the terminal, Cursor in the editor, v0 or Lovable for a quick UI, and ChatGPT when you want a second opinion. Four tools, four different ways to explain the same bug.

What voice alone still misses: "this one"

Dictation fixes the typing. It does not fix the pointing. Say "move this button under the field" to /voice and Claude Code gets a lovely clean sentence with no idea which button you meant. A settings page has forty of them.

The fix people actually want is the one you would use with a person sitting next to you: point at the thing and talk. That needs three pieces together:

  1. What you said, transcribed.
  2. What you pointed at, as a tight crop, not a full-screen screenshot with forty candidates.
  3. What it is called, the element's accessibility label and window title, so the agent can find it in the code.

That bundle is a brief. Hand it to any agent and the first attempt usually lands.

Why the tool you talk through should not care which model listens

Models leapfrog each other every few months. The best agent for your codebase in March is not the best one in September. If your voice and screen habit lives inside one lab's app, switching models means relearning how to explain things.

Deiko sits one layer below all of them. Double-tap Right Option, point, talk, and the brief lands wherever you drop it: Claude Code, Cursor, Codex and Windsurf get file paths, Claude.ai, ChatGPT and Gemini get images, and v0, Lovable and Replit get the crops pasted in. Same gesture every time. Change models on Monday and your hands do not notice.

And it stays on your Mac

The labs' versions run through their cloud, because that is where their model lives. Deiko does not have a model to protect, so it does not need your screen. Screenshots, on-screen text and accessibility text never reach Deiko's servers. Your voice goes out for transcription and is deleted once the brief exists. To file each brief in the right task, Deiko sends what you said, a one-line summary of it, your window titles (with anything that looks like a key stripped) and a few lines about your earlier tasks through its relay to a sorting model. Deiko keeps none of it. With no network, Apple's on-device recognizer takes over and nothing leaves at all. The privacy page spells out every byte.

tl;dr The labs agree you should talk and show instead of type. They disagree about which app you have to be in. Deiko's answer: whichever one you like this week.

Questions people ask

Does Claude Code have voice input?

Yes. Since March 2026 Claude Code has a /voice command: hold Space, speak, release, and the text lands in your prompt. It is dictation only; it does not capture your screen.

Can Cursor see my screen when I talk to it?

Not from the voice button. Cursor's mic dictates text into the Agent prompt. To show it a screenshot you paste or attach one separately.

What are ChatGPT Appshots?

A ChatGPT desktop feature that grabs the active window's image and exposed text into a chat. On a Mac you press both Command keys. It only sends to ChatGPT.

Is there a voice tool that works with every AI coding agent?

Deiko is built for exactly that: one gesture (point and talk) that produces a brief with cropped screenshots, which you can drop into Claude Code, Cursor, Codex, Windsurf, ChatGPT, Claude.ai, Gemini, v0, Lovable or Replit.

Does my screen get sent to the cloud?

Not with Deiko. Screenshots and on-screen text never reach Deiko's servers. Your spoken audio is sent for transcription and then deleted. To file each brief in the right task, Deiko sends what you said, a one-line summary of it, your window titles (with anything that looks like a key stripped) and a few lines about your earlier tasks through its relay to a sorting model. Deiko keeps none of it. Offline, nothing leaves the device.