backtalk

Never used Claude Code? Start at jaredrhod.com: pick your situation and it routes you to the right path.

Runs on: Claude Code only; the voice is built on Claude's agent SDK. The $20 Pro plan is enough.

Talk to your Claude Code agent out loud. Hold a key, say the thing, and it answers through your speakers in a real voice about a second later, with all its tools, your project context, and its own personality. Your AI finally has something to say back.

The hearing and the voice run local: free, offline models on your machine, no voice API keys, no per-word costs. The brain is the Claude Code you already have. On a Claude subscription, talking works like any other session and uses your plan's usage, with nothing extra to buy. This is the same voice loop I run every day, the one you see answering in about a second on my videos, shipped as working code so your agent's job is pointing it at your setup, not building it from scratch.

What it does

• Hold a key, talk, release. Your words are transcribed locally and handed to a live Claude Code session. The reply is spoken sentence by sentence as it's generated, with first audio in about 1 to 2 seconds on warm turns. Prefer no button at all? Hands-free listening is one spoken sentence away ("go hands free"), and the key keeps working there as your interrupt.
• It's YOUR agent talking. The session runs in the folder whose CLAUDE.md defines your assistant: same name, same personality, same memory as your terminal sessions. backtalk has no personality of its own; it's a mouth and ears for whoever you already have. (No agent yet? The ai-memory-vault build ships with mine, Jarvis, ready to use.)
• Interrupt it. Press the key while it's talking and it shuts up and listens. No headphones needed, because the mic only opens while you hold the key, so it never hears the speakers.
• Type instead whenever you want. Typing in the terminal is the same conversation, and the reply is still spoken.
• It asks before it acts, in plain words. When your agent wants to do something real, it asks out loud the way a person would ("I want to change a note in your vault called Recipes") and waits. An exact spoken yes approves; "details" reads you the exact command; anything else denies, with your words passed back as the reason, so "no, put that in drafts instead" actually steers it. Most read-only work passes without interrupting. Prefer auto-approve? Say "stop asking for permission" (or "turn off the permission prompts") and confirm; it changes its own config, and the first ask of every session reminds you the phrase exists.
• The voice console. Session control by voice, so you never go back to the keyboard: "clear the session", "compact the session", "switch to the deep model" / "back to the fast model", "set effort to low" (or medium, high, max; this one saves itself as your default), "usage report", "go hands free" / "push to talk mode" for the microphone, "stop asking for permission" / "start asking again" for approvals. Exact phrases, spoken alone. (Credit where due: this grew out of a community member's own build shared in the Discord.)
• It can pick up where it left off. Set "resume_last_session": true in the config and every launch reattaches to your previous conversation instead of starting cold, so closing the window stops costing you the thread. Off by default. And the built-in voice has a pace dial: "speed" in the config, 1.0 native, 1.15 brisker. (Credit where due: both grew out of a community proposal by aram-cloudstak.)
• Music ducks while it speaks (Spotify, macOS) and comes back up after.
• It thinks out loud. While the agent works, you hear the processing sound from my videos, so a pause never reads as a dead line. Silence it with "thinking_sound": "" in the config.

Install

git clone https://github.com/jaredrhod/backtalk
cd backtalk
./install.sh

The installer sets up a Python environment, the two local AI models (speech-to-text and the voice), and the one system library they need. First run downloads the models (about 1 GB total); everything after is instant. Prerequisites: Claude Code with a Claude subscription, and uv (the installer offers to install it).

The easy way to configure it: open this folder in Claude Code and say "read backtalk.md and set me up." The wizard picks your agent folder, your key, and your voice with you, then test-fires the whole loop.

Already in a Claude Code session with your agent? One sentence does the whole install: "clone https://github.com/jaredrhod/backtalk.git, then read backtalk/backtalk.md and set me up." Your agent runs the installer and the wizard for you.

The manual way: copy backtalk.json.example to backtalk.json (your copy is untracked, so updates never touch it), then edit it. Point agent_dir at the folder whose CLAUDE.md is your agent, set name to your agent's name, pick a ptt_key. Then:

./run.sh

Hold the key. Talk. Let go.

Windows

Windows is the newest lane, and the setup runs through the wizard instead of the shell scripts (install.sh and run.sh are Mac and Linux). Open this folder in Claude Code and say "read backtalk.md and set me up": the wizard installs uv, espeak-ng, the environment, and the models natively, then launches with uv run python -m backtalk.main. The ElevenLabs key lives in the ELEVENLABS_API_KEY environment variable on Windows for now (Credential Manager support is planned). Hit something rough? The Windows notes in TROUBLESHOOTING.md carry the known quirks, and issues are welcome.

The voice

Two engines, and the setup wizard offers you both instead of quietly defaulting.

Built-in (Kokoro), the free one. Local, offline, no accounts, no per-word costs, and honestly a bit computer-sounding. The default voice is bm_lewis, a British male with exactly the butler register. Around 60 voices ship free; set voice in backtalk.json (the first letter picks the language: a is American, b is British, and there are Spanish, French, Hindi, Italian, Japanese, Portuguese, and Chinese voices too).

ElevenLabs, the natural one. The human-sounding voice most people actually want, on your own API key. The free tier is enough to audition it; day-to-day talking runs on the paid starter plan. The wizard walks the whole thing with you: account, key into the keychain, then an audition of real voices through backtalk's own mouth until one fits. Want the exact voice from my videos? It's called Tarquin in the ElevenLabs voice library: search it by name and you're done hunting. Under the hood it is: set elevenlabs.enabled and your voice_id in the config, and have ffmpeg installed. The key never goes in a file. On macOS, seed it into the Keychain once with security add-generic-password -a "$USER" -s backtalk-elevenlabs -T /usr/bin/security -w (it prompts for the secret) and backtalk reads it from there. Linux: secret-tool store --label backtalk service backtalk-elevenlabs. The ELEVENLABS_API_KEY environment variable works as a last resort, but an export in a shell profile is a plaintext key on disk; the keychain is the grown-up path. Kokoro stays wired in as the automatic fallback, so if the cloud fails the voice degrades instead of going mute, and logs/backtalk.log records why.

Give it a face (optional)

backtalk writes tiny state files while it listens, thinks, and speaks, so anything can watch them and react in real time.

• ai-visualizer is the matching face: four full-screen visualizers, including the living circuit board from my videos. Point its bus_dir at this folder (or set signals_dir here to its folder) and it performs your actual conversation, idling, listening, thinking, and speaking along with the voice.
• barehands: point barehands_state_dir at its state/ folder and the on-screen ring becomes your agent's face, breathing while idle, spinning while thinking, and pulsing with the voice while it talks.

Mind (ai-memory-vault), mouth (this), face (ai-visualizer), hands (barehands).

The fine print that matters

• Usage: every spoken turn is a real Claude Code turn, so a long voice session uses your plan the same way a long typing session does. The config pins the fast model tier on purpose; it's most of the speed, and it's the lighter draw.
• Permissions: ask first, auto-approve by choice. The default is "ask": gated actions get a spoken permission check, answered by voice or by typing, and silence for about 75 seconds means no. "bypassPermissions" is auto-approve: the agent acts without asking, exactly like a terminal session with approvals off. Never hand-edit the file to switch; tell your agent to change it (takes effect the next time the voice line starts), or say "stop asking for permission" / "start asking again" in a voice session for a flip that happens immediately and saves itself.
• Two microphone modes, and the words mean what you think. Push to talk (the default): the mic is closed except while you hold the key, so nothing records in the background, ever. Hands-free listening: always listening with voice detection; the setup asks which you want, "go hands free" / "push to talk mode" switches live and saves itself, and the talk key still works in hands-free as your interrupt. Tradeoffs in TROUBLESHOOTING.md. (Hands-free is about the MICROPHONE. Approvals are a separate setting called auto-approve; the two never share a name.)
• The talk key needs a global key listener. For the key to work when the voice line isn't your focused window, the process has to watch keyboard events system-wide. backtalk/ptt.py compares each event against the one key you