Fair things to ask.

The ones that come up before people download it, answered with the mechanism rather than the reassurance.

Does anything I say get uploaded?

No, and not as a policy: as an absence. Recognition and synthesis both run on your Mac against models in your own home directory, so there is no code path that sends audio anywhere. The app makes three kinds of request on its own: fetching those models the first time you need them, asking whether a newer build exists, and, once you have paid, claiming a seat for this Mac. That last one sends the licence key, a hash of the hardware id and the name you gave the computer. None of the three carries a word you said, and the privacy page spells all of it out.

Which agents does it work with?

Claude Code, Claude Desktop, Codex, Cursor, opencode, Windsurf, Zed, Cline, Gemini CLI and Hermes are detected and set up for you. Anything else that speaks MCP can use it too: setup gives you the block to paste, and it will offer to write into a config it finds that it does not recognise. There is nothing agent-specific in the protocol.

How is this different from dictation that is already on my Mac?

Dictation is one direction. It turns your speech into text and stops there. The thing Earshot adds is the other direction: an agent can ask you a question out loud and hold its turn open until you answer, so a session stops being a queue of prompts. The dictation is there too, and it is good, but it is not the part that is new.

Will it understand my accent?

Usually, and where it does not, the fix is a text file. Every recogniser mangles a particular handful of words for a particular person, and vocabulary rules are one per line and applied after recognition, so a rule you got wrong is visible on screen instead of buried in the model.

Can I use it in a room with other people?

Push to talk is the honest answer: it hears nothing until you hold the key. Hands-free listens for a wake word, and in a noisy room you will want to raise the wake threshold, which trades a little responsiveness for not transcribing the person next to you.

What happens when it mishears an answer to a yes-or-no question?

It hands the agent your words rather than its best guess. Where an agent supplies a list of choices, your answer is matched against it, and anything matching none of them comes back verbatim. An agent given the wrong option confidently has no way to notice, so the design errs towards passing through what you said.

Do I need a licence key to try it?

No. Free is one person and one agent, with nothing switched off and no clock running: dictation, all three tools, every voice, the hotkey, and vocabulary. Pro starts to matter when there is more than one agent, session or machine in play.

What happens to my key if you disappear?

It keeps working for about six weeks and then it needs its seat renewed, which is a request to this website. Day to day nothing asks us for permission: the check is a signature your Mac verifies by itself, which is why there is no account to sign into and why a plane or an outage costs you nothing. But the seat behind it is renewed here monthly, and if we were gone there would eventually be nothing to renew against. Ask us before you buy if that matters to you.

Does it work on an Intel Mac?

Not today. Both models run locally and they lean on Apple silicon to start answering in under a second, which an Intel machine cannot do well enough to be worth shipping.

How much disk does it take?

About 800 MB, almost all of it models, kept outside the app bundle in ~/.cache/earshot. Updates move roughly forty megabytes because the models stay where they are.

Something not answered here? Ask us directly.

Wanting models of your own instead? Book twenty minutes.