Real-time, offline text-to-speech for AI coding agents — no API keys, no cloud, no lag.

Your AI agent's responses are spoken aloud as they complete. Hands-free. Runs locally. Works everywhere.

One service. Any agent. Any IDE.

pip install git+https://github.com/omernesh/AgentTalk
agenttalk setup

Runs on: Windows · macOS · Linux · Python 3.11+

For specific integrations, pass additional flags to setup:

agenttalk setup                   # Claude Code (default)
agenttalk setup --antigravity     # + Google Antigravity IDE
agenttalk setup --opencode        # + opencode

After setup, start the service:

python -m agenttalk.service       # foreground
pythonw -m agenttalk.service      # background (Windows)

Or double-click the AgentTalk desktop shortcut created by setup.

AgentTalk runs a local HTTP service on localhost:5050. Agent hooks and extensions call POST /speak at the end of each response. The service synthesizes speech using one of two offline engines — no cloud, no API keys.

Audio ducking (Windows): While AgentTalk is speaking, all other system audio is automatically lowered to 50% volume via the Windows Core Audio API. Volume is restored the moment speech finishes — so you never miss a word over music or a video.

Claude Code / VSCode / Antigravity / opencode
           ↓
    POST localhost:5050/speak
           ↓
    Kokoro / Piper
           ↓
      🔊 Your speakers

Type these as your message in Claude Code (or any supported agent):

30 voices across four accent families, all local, all offline:

Switch anytime with /agenttalk:voice [name] or the tray menu. Changes take effect on the next utterance and persist across restarts.

Right-click the system tray icon:

• Mute / Unmute — instant toggle, checkmark shows current state
• Model — submenu: switch between Kokoro and Piper (radio buttons)
• Voice — context-aware submenu: Kokoro voices when on Kokoro, Piper .onnx stems when on Piper
• Active: {voice} — read-only display of the currently selected voice
• Quit — stops the service and removes the tray icon

The icon animates while speaking and returns to default when playback finishes.

All settings live in the platform config directory and persist across restarts. Change them at runtime — no restart needed.

agenttalk setup automatically registers Stop and SessionStart hooks in ~/.claude/settings.json.

agenttalk setup --antigravity

Installs a native Antigravity skill to ~/.gemini/antigravity/skills/agenttalk.md and a session startup workflow to ~/.gemini/antigravity/global_workflows/agenttalk_start.md. Alternatively, install the VSCode VSIX extension directly — Antigravity is a VS Code fork and the extension is fully compatible.

Install the extension from integrations/vscode/:

code --install-extension integrations/vscode/agenttalk-vscode-1.0.0.vsix

The extension auto-detects Roo Code and KiloCode and hooks their message events. A status bar item shows service state.

agenttalk setup --opencode

Registers session_start_hook.py and stop_hook.py in ~/.config/opencode/hooks/.

# Pipe any CLI tool through AgentTalk
openai api chat.completions.create … | python integrations/openai-cli/stream_speak.py

Or source the shell function from integrations/openai-cli/README.md for a speak-openai shortcut.

The skill file at integrations/openclaw/SKILL.md (also available at the URL below) contains both the installation walkthrough and the agent operating instructions. Point OpenClaw at it and ask it to install AgentTalk — the agent follows the steps autonomously.

Option A — let OpenClaw install from the raw file:

In an OpenClaw session, paste this prompt:

Read https://raw.githubusercontent.com/omernesh/AgentTalk/main/integrations/openclaw/SKILL.md
and follow the AI Agent Installation Instructions to install AgentTalk on this machine.

OpenClaw will run pip install agenttalk, call agenttalk setup --no-autostart, start the service, and verify it is healthy — all without leaving the session.

Option B — register as a persistent skill:

Add the skill to your ClawHub workspace so it loads automatically every session:

clawhub install https://raw.githubusercontent.com/omernesh/AgentTalk/main/integrations/openclaw/SKILL.md

Once installed, OpenClaw checks the service is running at session start and speaks each response automatically (auto mode) or on demand via /agenttalk:speak (semi-auto mode).

The service exposes a simple HTTP API at localhost:5050:

# Speak text
curl -s -X POST localhost:5050/speak \
  -H "Content-Type: application/json" \
  -d '{"text": "Hello from AgentTalk"}'

# Check service status
curl localhost:5050/health

# Get / update config
curl localhost:5050/config
curl -X POST localhost:5050/config -d '{"voice": "bf_emma", "speed": 1.2}'

# Mute / unmute
curl -X POST localhost:5050/mute
curl -X POST localhost:5050/unmute

1. Check it's running: curl localhost:5050/health
2. Check hooks are registered: look for agenttalk in ~/.claude/settings.json under hooks.Stop and hooks.SessionStart
3. Re-run agenttalk setup to re-register (idempotent — won't duplicate)

Symptom: No audio, or PaErrorCode -9984 in the log.

Fix: In Windows Sound Settings, set your output device to "Shared" mode. Check config.json dir for agenttalk.log to see which audio mode was detected.

Re-run agenttalk setup — downloads are idempotent. Ensure github.com is reachable and you have ~400 MB free.

The tray icon uses pystray, which has a known GIL compatibility issue on Python 3.12 on Windows. Use Python 3.11 if you need the tray icon on Windows. macOS and Linux are unaffected.

py -3.11 -m venv .venv && .venv\Scripts\activate
pip install git+https://github.com/omernesh/AgentTalk
agenttalk setup

MIT