Introducing Loudkit for agents

Some thoughts arrive before you can type them neatly. A question while you make coffee. An idea you want to say out loud. A long answer you would rather listen to. Voice notes make room for those moments without asking you to keep looking at a screen.

We want talking to your agent to feel as easy as sending a message to a buddy. You keep the agent you chose, with the context and tools you have already given it. Loudkit supplies the speech layer. Your agent still decides how to answer and which actions need approval.

This follows the idea behind LoudReader: make useful things easier to listen to. Our Loudkit framework announcement covers the speech toolkit for developers. This companion puts it to work in conversations with an existing agent.

With a compatible connection configured, the companion transcribes incoming audio with Parakeet, passes the text to your agent, and turns its reply into speech with Loudkit. The messenger connection is responsible for returning that audio to the right conversation.

Speech generation and transcription run on your Mac after the models are downloaded. Your agent's language model and the messenger may still use cloud services. Local speech does not make the whole conversation offline, or change those services' privacy policies.

For Hermes and OpenClaw, the preferred route is to configure their speech providers while keeping the existing bot, sessions, memory, tools and permissions. Other agents need a suitable CLI, API or active inbox integration. Compatibility depends on the actual interface your agent exposes.

Real audio has been generated and transcribed locally. We have also tested pinned Hermes and OpenClaw speech-provider code against the local service. Those checks establish that the speech pieces can talk to one another. They do not establish that a complete bot has received and answered a voice note through a live messenger account.

The repository includes adapters for Telegram, Discord, WhatsApp Business Cloud API, Slack and BlueBubbles/iMessage. They have protocol tests; real account delivery remains to be checked. WhatsApp uses its Business API, not a personal-account QR login. See the published verification scope before choosing a connection.

1. Listen first. The Loudkit for agents page has voice samples and a ready-to-copy setup prompt.
2. Give your agent the guide. Share the agent setup instructions and ask it to check your agent version, messenger and speech host. The current runtime needs an Apple Silicon Mac, Python 3.12–3.13, uv, FFmpeg and at least 20 GB of free disk space. The initial model download is about 3 GB.
3. Verify one small conversation. Start with a working text reply, then local speech. Finally, try a harmless voice note in your chosen chat and confirm that the reply arrives and plays. Treat that last step as part of setup.

OpenClaw users can also install the Loudkit Voice setup skill on ClawHub (version 0.1.1). It helps the agent follow this process; installing instructions does not install or start the speech runtime.