Loudkit voice notes — developer preview

Help the user speak to the agent they already use and receive a spoken reply in the same chat. Loudkit supplies speech synthesis; the free, open-source Loudkit Voice companion adds Parakeet transcription and connection options. Its service and command are currently called LoudTalk. Installing this skill adds setup instructions; running the companion and downloading models are separate setup steps.

Read the current agent setup guide before configuring a connection. It is the authoritative source for provider examples, tested revisions and preview availability. The product page has voice samples and the human explanation. If the guide is unavailable, give a compatibility assessment from the constraints below; do not guess release-specific configuration fields.

Assess the existing setup

Identify the agent and installed version, messenger, agent host, and whether an ordinary text conversation already works. Inspect available local information first; ask only for missing facts. Determine whether the companion is already running or needs to be cloned from its public repository. An installed Loudkit SDK alone is not the voice-note companion.

The current companion requires an Apple Silicon Mac, Python 3.12–3.13, uv and FFmpeg. Allow about 3 GB for models and keep at least 20 GB of disk free during setup. Check these on the intended speech host with read-only commands such as uname -sm, the intended Python interpreter's --version, command -v uv, command -v ffmpeg and df -h. A different current agent host can still be assessed, but needs a deliberately configured authenticated connection to the Mac; its own 127.0.0.1 does not point to the user's Mac.

The source is available without a signup, subscription or access request. If the intended speech host lacks a requirement, report that concrete blocker and continue any useful compatibility assessment. pip install loudkit installs the speech SDK, not LoudTalk. The user's agent or messenger provider may have its own costs. Do not open a public issue on the user's behalf unless asked.

Choose the connection

Keep the existing agent's sessions, memory, tools, model credentials, messenger connection and approval rules. A generic model API is not a connection to that agent's existing session. An MCP inbox needs an active agent to read and reply; it does not wake a stopped agent. This skill does not expose a hosted MCP server.

Run the companion and configure speech

Install missing prerequisites on the intended speech host within the user's setup request. With Homebrew, brew install uv ffmpeg supplies those two tools. Clone and run the companion:

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

The launcher installs locked dependencies, including Discord support, then opens http://127.0.0.1:8765. Prepare the speech models in that local setup page. A separate Loudkit checkout is not required. For updates and command-line options, use the source README.

For Hermes and OpenClaw, read the current official Hermes voice documentation or OpenClaw audio and TTS configuration for the installed release, then merge only the required speech fields from the Loudkit guide into the active profile.

The preview speech API defaults to http://127.0.0.1:8765/v1, with transcription model parakeet, speech model loudkit and English voice sophie. The local SDK placeholder loudtalk-local belongs to speech configuration only; it must not replace the agent's conversation-model credentials. Preserve a recoverable copy of any configuration you change. Use the existing gateway as the sole receiver of its bot's messages; do not start another Telegram poller with the same token.

Optional standalone messenger bridges are protocol-tested, not verified on real accounts: Telegram, Discord, WhatsApp Business Cloud API, Slack and BlueBubbles/iMessage. WhatsApp does not provide a personal-account QR login; Slack uses audio attachments. Select a bridge only when it fits the user's existing agent and requested messenger. Keep credentials in local configuration, not chat or public issues, and preserve the agent's enabled tools and permissions.

Verify the actual journey

First confirm an ordinary text response. Generate a short sample and transcribe a clear recording locally, then send a harmless voice note such as “Tell me a short joke” through the user's messenger. Confirm that the reply arrives in the same chat and plays. If it fails, isolate transcription, agent dispatch, speech synthesis and message delivery; avoid repeated account or configuration changes without a new diagnosis.

Report the exact agent version, messenger and observed result. Distinguish compatibility assessed, locally verified speech, and a verified messenger round trip. Speech runs locally after model download; the agent and messenger may still use online services.