FAQ
Short answers to common questions about platforms, cost, privacy and what the brain may do.
Short answers, each with a link to the page that covers it in full. For something that isn't working, see Troubleshooting.
No. The echo canceller is Apple's voice-processing unit.
claudio makes two choices, each with its own platforms:
- Where your sessions live, the
backendsetting: cmux, whose panes are the sessions, or bb, whose threads are. A session can run any agent the platform offers, though only Claude Code sessions report a finished turn. - Where you see and type to claudio, the front end: a terminal, or bb's panel, which the bb plugin adds beside every thread.
It's the same claudio everywhere, with the same settings and rules. No editor front end exists yet, but nothing in claudio is tied to one: any front end can connect to claudio's view.sock, read what it says and type to it. Front ends describes the protocol, and Platforms compares the choices.
The brain runs through your Claude Code login, so it uses your Claude plan like any other session. Deepgram bills the listening and, with the default mouth, the speaking. ElevenLabs bills the speaking if you switch the mouth to it.
Not unless you allow it. Out of the box the brain can read and search files, run git status, git log, git diff and git branch (to list), read GitHub with gh (issues, pull requests, search), use a few commands of your sessions' platform (cmux or bb) to list, read and open sessions and to answer a session's question with the option you chose, and run claudio set. Work goes to your sessions, which run with their own permissions. Anything more is up to you: say yes or always when it asks, or list it in settings.json. See what the brain may do.
- Permission prompts: only with your answer. It reads the prompt out, then asks you yes, always or no; claudio's own code, never the brain, gives the session your answer: in cmux it types it into the session's pane, after checking the prompt is still there; in bb it answers through bb. Nothing a session prints can answer it.
- Multiple-choice questions: yes. It reads the options aloud, you say which one, it confirms, then answers it with the option you picked: in cmux with
claudio choose, which refuses anything but a question and its own options; in bb withbb thread interactions answer.
Your CLAUDE.md, yes. Your settings, no.
- CLAUDE.md: the brain reads the CLAUDE.md files that Claude Code loads for the project you run claudio in, like any session. Keep instructions there that you are happy for the voice to follow too.
- Settings: the brain starts with
--setting-sources "", so yoursettings.jsonfiles do not apply to it: not your permissions, hooks or env. Which tools it may use is set by claudio, plus what you allow in~/.config/claudio/settings.json, and it can't read~/.config/claudio/secrets.json. - MCP servers: none of yours, only those you list in
~/.config/claudio/mcp.json. - Skills and slash commands: none. Its prompt lists claudio's own slash commands, so it can tell you about them.
Your sessions use your settings as usual.
Your audio goes to Deepgram to be transcribed. What you say goes to Claude, as the brain's turns. Replies go to Deepgram, or ElevenLabs, to be spoken. Transcripts and logs stay on your Mac, under ~/.config/claudio.
Set ears to flux-general-multi, edit the persona so the brain answers in your language, and pick a voice that speaks it. See other languages.
Turn on address_me with claudio set address_me true. Then only turns that start with its name count ("claudio", or one you give it with "slash name Roger"), and the rest show dim in the transcript. Say the name alone and the next turn counts without it.
Put a .claudio.json in the project and run claudio from there. Any setting except the API keys can go in it. See per-project settings.
The persona asks the brain to say what it's about to do before each tool call. When the brain skips that, claudio fills in one word, "Looking." before it reads or searches files and "Checking." before anything else, so you don't sit through silence.
Answer the question. When the brain needs a command its rules don't cover, the command waits and claudio asks, in the same reply, “May I run gh pr create?”, with yes, always and no listed under it; say one, or press 1, 2 or 3. Yes runs it once; always also adds the rule to ~/.config/claudio/settings.json, for every later run; no, or a minute's silence, doesn't run it.
You can also edit that file yourself, in Claude Code's settings format, such as {"permissions": {"allow": ["Bash(gh:*)", "WebFetch"]}}; saving it restarts the brain. Telling the brain “I approve” does nothing on its own: only claudio's question can allow it. See allow more.
No. With memory on, the brain resumes the same conversation the next time you start claudio, if it's less than 7 days old, so Ctrl-C and a restart keep everything it said and believed. That includes a wrong conclusion, such as "the session tools are down" after an outage that has passed. To start fresh:
- say "slash clear", or type
/clear, while it runs - run
claudio forgetbefore the next start - run
claudio set memory falseto start fresh every time
None of these delete your transcripts. See what the brain remembers.
Remove the hooks first, then the tool, then its files:
claudio uninstall
uv tool uninstall claudio
rm -rf ~/.config/claudioThe last line deletes your settings, keys, persona and transcripts.