Voice notes for your agent

Loudkit for agents is a developer preview: local speech synthesis with Loudkit, transcription with Parakeet, and audio replies through an existing bot or a messenger bridge. It does not replace the agent that handles your requests.

The companion is free and open source in the Loudkit Voice repository. Its local service and command are currently called LoudTalk. You can run it from source; there is no signup or access request. pip install loudkit installs the speech SDK, not this companion. Start by checking your agent, its version, messenger and operating system against the routes below.

Choose a connection

Your setup	Route	Verified scope
Existing Hermes or OpenClaw bot	Configure its speech providers to use LoudTalk	Pinned upstream speech-provider code tested against real local speech models
Local Codex	Voice transcription, CLI request, spoken reply	Real local agent round trip; messenger transport was mocked
Grok Bot	Experimental skill reads and replies to a local inbox	Local transport tested; actual Grok Bot execution pending
Muse Code	CLI or MCP adapter	Configuration recipe; runtime validation pending
Other CLI, API or MCP agent	An explicit adapter or agent-managed inbox	Depends on that agent; protocol support alone is not a complete integration

The provider tests do not load a complete Hermes/OpenClaw gateway or prove delivery to real messaging accounts. Configuration may differ between releases. Hermes already offers voice replies. OpenClaw already supports audio transcription, including local Parakeet. This integration supplies a local speech service and a choice of Loudkit voices.

Before connecting a bot

The current companion requires an Apple Silicon Mac, Python 3.12–3.13, uv and FFmpeg. Models download on first use; allow approximately 3 GB for them and keep at least 20 GB free during setup. The agent must already work with text messages.

Run the companion

On the Mac that will run speech, install uv and FFmpeg if they are missing. With Homebrew:

brew install uv ffmpeg

Clone the free, open-source companion and start it:

git clone https://github.com/loudreader/loudkit-voice.git
cd loudkit-voice
./start.command

The launcher installs the locked dependencies, including Discord support, and opens the local setup page at http://127.0.0.1:8765. Prepare the speech models there before connecting your agent. The default speech API is http://127.0.0.1:8765/v1. A separate Loudkit checkout is not required.

For source updates and command-line options, follow the repository README.

127.0.0.1 means the machine running your agent. An agent hosted elsewhere needs a deliberately configured, authenticated connection or a tunnel to the Mac. The server binds to loopback by default. The website at loudkit.loudreader.io is documentation and audio samples; it does not run the speech models or connect to your messenger account.

Existing Hermes bot

Merge these fields into the active Hermes profile; preserve the rest of its configuration. This example selects the English voice Sophie.

stt:
  provider: openai
  use_gateway: false
  openai:
    base_url: http://127.0.0.1:8765/v1
    api_key: loudtalk-local
    model: parakeet
    timeout: 300
    max_retries: 0
tts:
  provider: openai
  use_gateway: false
  openai:
    base_url: http://127.0.0.1:8765/v1
    api_key: loudtalk-local
    model: loudkit
    voice: sophie
    speed: 1.0

loudtalk-local is a placeholder required by the SDK for local audio, not an OpenAI credential. Do not replace the credentials used by your conversation model. Restart the existing gateway, then use /voice on in its Telegram or Discord chat. /voice status shows the saved setting. Other messenger behaviour depends on the Hermes release and channel.

Keep the existing bot as the sole receiver of its messages. Do not give the same Telegram bot token to a second poller.

Existing OpenClaw bot

In the preview setup page, generate an OpenClaw speech configuration and merge it into the existing installation. Select a separate audio authentication profile so local speech credentials do not replace the credentials for your main model.

The provider API uses these values:

Setting	Value
Base URL	http://127.0.0.1:8765/v1
Transcription model	parakeet
Speech model	loudkit
Example English voice	sophie
Local SDK placeholder key	loudtalk-local
Automatic TTS behaviour	inbound — speak replies to incoming audio

Check the field names for your installed release against the current OpenClaw audio guide and TTS configuration. The tested provider revision is b068eab40767fc6796899cdc85d08102e8035b2b; full gateway configuration loading was outside that test.

OpenClaw’s messenger support evolves independently. In particular, do not infer current native iMessage support from an older BlueBubbles example. The preview’s standalone BlueBubbles bridge is a separate connection.

Other agents and messenger bridges

The preview includes protocol-tested bridges for Telegram, Discord, WhatsApp Business Cloud API, Slack and BlueBubbles/iMessage. Live delivery on real accounts still needs validation. WhatsApp is not a personal-account QR login; Slack replies use audio attachments.

A CLI/API adapter can submit a transcribed request and return an answer. An MCP inbox instead waits for an already active agent to read the message and reply. MCP does not wake a stopped agent. Connecting to an LLM API does not connect to a separate agent’s existing memory, tools or session.

What runs locally

Speech synthesis and transcription run on the Mac after model download. The messenger, agent and language model can still use online services. The entire conversation is not promised to be offline.

Transcription can make mistakes, especially on short or ambiguous recordings. Keep the agent’s normal permissions and approval rules for consequential actions. The current preview accepts recordings up to 3 minutes and 25 MB.

Test your own setup

1. Confirm that the existing agent responds to an ordinary text message.
2. Play an English voice sample locally and transcribe a clear sentence.
3. Send a harmless voice note to the connected bot, such as “Tell me a short joke.”
4. Confirm that the reply arrives in the same chat and that the audio plays.
5. Record the exact agent version, messenger and result when reporting feedback.

Local provider and protocol tests are a starting point. A successful voice note on your own account is the test that completes the connection.