agentic-tooling-wiki

Claude-to-Speech

A community Claude Code plugin that adds automatic text-to-speech to Claude’s responses via the ElevenLabs API. A small, focused example of the Claude Code plugin ecosystem — agent I/O tooling rather than a capability/knowledge skill.

How it works

1. User activates voice mode with the /speak slash command.
2. Claude embeds invisible TTS markers in its responses (e.g. <!-- TTS: "…" -->).
3. A Stop hook extracts those markers from the transcript.
4. claude_speak.py sends the text to ElevenLabs (or a local TTS server).
5. Audio plays through the system’s default player.

Features

• Smart defaults: silent for code dumps (<!-- TTS: SILENT -->), vocal for questions/confirmations.
• Dual-mode: direct ElevenLabs API or a local TTS server.
• Multiple preset voices + custom voice IDs; 2-second dedup; macOS/Linux (incl. Raspberry Pi)/Windows.

Requirements & status

ElevenLabs API key, Python 3.7+, requests. MIT, ~89% Python. Early-stage community project: 15 stars, 6 forks, 19 commits.

Where it fits

A concrete instance of the agent-skills idea — capability packaged as files an agent loads — but on the output/UX side (voice) rather than reasoning. Sibling to compound-engineering-plugin as another Claude Code plugin, and runs on the same hook/slash-command surface that hosts like claude-cowork expose. A lightweight agent I/O facet of the ecosystem; notable mainly as evidence the Claude Code plugin surface is attracting small third-party tools. Arrived via Telegram, routed here by the hub router (../HUB.md).