voice-access

Use a Windows computer by voice, with Claude.

An open-source project for people who want to talk to their computer and hear it answer. During setup, it checks with you, out loud, that you really heard it.

Get the code and setup instructions on GitHub Is it for me?

What is this?

voice-access has two parts. One is for talking with Claude. The other gets your computer's sound working for you first.

talk: a spoken conversation with Claude

  1. You speak.

    Your voice stays on your computer.

  2. Your words are written out.

    This happens on your own computer.

  3. Claude Code answers.

    Claude receives the text, not the audio.

  4. The answer is read aloud.

    Then it is your turn again.

Say things like

  • "Repeat that"
  • "New conversation"
  • "Permissions"
  • "Goodbye"

Beeps tell you when it is

  • Listening
  • Working
  • Stuck on a problem

accessible-voice-setup: sound that really works for you

A Claude skill, which is a set of instructions Claude can follow. It finds out what your computer's sound input and output can really do. It confirms each one with you using a repeat-back test, then builds one-word voice commands around what works.

If your computer is missing something, it can help you compare and price a microphone, speaker, headphones or memory.

The idea behind it

A successful API call is not evidence that a human heard anything. Confirm with the person.

In plain words: a computer reporting that it played a sound does not prove you heard it. So voice-access asks you.

Spoken approvals

A voice session has no screen to click "allow" on. So talk reads each request aloud, including the actual command. You say yes or no, or ask "what does that do?".

A very long command is cut off after about 200 characters when read aloud. The full request appears on screen and in the log.

  • A short, clear yes is the only thing that counts as yes.
  • Silence means no.
  • An unclear answer means no.
An example of how it can go. Claude's descriptions and explanations will vary.
  1. talk says: "Claude wants to run a command it describes as: delete the old notes file. The actual command is: del old-notes.txt. Say yes to allow it, no to refuse, or ask what does that do."
  2. You say: "What does that do?"
  3. Claude says: "Nothing has run yet. It would delete the file called old-notes.txt. Do you want me to do that?"
  4. You say: "No."

Is it for me?

It may suit you if

  • Typing or using a mouse is hard, tiring, painful or impossible for you.
  • You use Windows and would like to talk with Claude and hear the answers.
  • You are helping someone else get set up.

It may not suit you if

  • Your computer does not run Windows.
  • Speaking, or hearing spoken answers, is not an option for you. talk is a spoken conversation.
  • You cannot type and have nobody to help with first-time setup.

So far, it has been tested by voice on one Windows computer, so you may run into problems nobody has seen yet.

What do I need?

  • Windows
  • Python
  • Claude Code
  • A Claude account

Python is a free programming language that talk runs on. Claude Code is Anthropic's program for using Claude from a terminal, which is a window where you type commands.

And someone who can type, if you cannot. First-time setup uses a few terminal commands, so you may need a helper for that part.

How do I get started?

  1. Read what it can access, so you know what you are agreeing to.
  2. Open the setup instructions for talk on GitHub. Everything you need to install is listed there. To add the skill, see the project's main page.
  3. Run the terminal commands from the instructions, or ask your helper to.
  4. Start talking. It starts with the smallest set of permissions, and you can choose how much to allow.

If you are helping someone

You may only be needed for the typing. Go through what it can access together first, and let the person decide. Then check that it works by asking them, not by reading the screen.

What can it access?

What leaves your computer

No audio leaves your computer. Claude receives the transcribed text, along with any files it reads and anything else it works with.

What it can do without asking

It starts with the smallest set of permissions. Without asking you first, it can only:

  • Read and search files in the working folder, which is the folder you start it from
  • Run commands there that only read information, such as git status

Any allow rules you have set in your own Claude Code settings also run without asking.

For anything else, it asks out loud first. That includes changing a file, reading a file outside the folder, and using the web. See spoken approvals.

Asking is not the same as blocking. If you say yes, Claude can go outside the folder for that one action.

You decide how much to allow

You can stop being asked about some things, and change your mind later, by voice. Say any of these on its own after the beep.

  • "Permissions" tells you what it can do now, and your options.
  • "Allow editing" lets it create and change files in the working folder without asking. It could overwrite your work.
  • "Allow reading" lets it read and search files anywhere on your computer without asking. Anything it reads is sent to Claude.
  • "Allow web" lets it search the web and open web pages without asking. A harmful page could try to trick it.
  • "Ask before editing", "Ask before reading" or "Ask before web" goes back to asking.

Before allowing anything, it reads you the risk and waits for a short, clear yes. Going back to asking needs no yes.

Allowing both reading and web is the widest choice, because a harmful page could try to trick Claude into sending a file's contents somewhere.

Your choices are saved for next time, and apply in whichever folder you start it from. Claude cannot change them for you. A helper can also change them by typing.

What it keeps

Unless you turn this off, it keeps a full written log of each conversation on your computer.