Lets Claude Code add Inworld voice features to your app: text-to-speech, speech-to-text, voice cloning and design, and realtime voice agents.
Built for Claude Code (terminal, IDE extensions, or the desktop app's Code tab). The voice tools run on your machine as a local MCP server, so they don't load in claude.ai chat or Cowork.
Once it's installed, ask Claude Code for things like:
- "Add a read-aloud button to our blog posts." Claude previews voices with you, then wires up the API route and the button.
- "Transcribe our support-call recordings with speaker labels." Claude writes a batch job using Inworld STT.
- "Build a voice agent for our booking flow." Claude scaffolds a realtime voice client with secure browser auth.
- "Design a voice for our game's narrator." Claude generates previews you can listen to, then saves the one you pick.
- "Why does my TTS call return 401?" Claude checks Inworld's docs and your code, then fixes it.
Without the plugin, Claude has to guess at Inworld's APIs. With it, Claude calls Inworld directly while you work, so you hear real samples before anything is wired in, and the code it writes follows Inworld's documented patterns.
/plugin marketplace add inworld-ai/claude-code-plugin
/plugin install inworld
Claude Code will prompt for your Inworld API key. Get one at platform.inworld.ai/api-keys — paste the Base64-encoded value as-is.
| Command | What it does |
|---|---|
/inworld:setup [tts|stt|both] |
Detects your stack (Next.js, Express, FastAPI, Flask, plain Node/Python, etc.) and scaffolds a lib/inworld.{ts,py} helper module with TTS-2 + STT clients, adds INWORLD_API_KEY to .env.example, picks a voice, runs a Hello-from-Inworld test. |
/inworld:add-voice [where] |
Wires voice output into one specific feature. Browses voices, demos them with and without steering, then edits the relevant route or component. |
/inworld:add-realtime [browser|server|twilio] |
Scaffolds a working Realtime client — WebSocket for server / Twilio, WebRTC for browser. Includes backend token minting, mic capture, audio playback, session.update wiring. Includes an OpenAI Realtime migration cheat sheet if you're already on that. |
/inworld:add-streaming-tts [http|websocket] |
Scaffolds a streaming TTS client for sub-200ms time-to-first-audio. Chunked HTTP for simplicity, WebSocket for cancellation + barge-in support. |
/inworld:add-transcription [short|long] [diarization] |
Scaffolds transcription of recorded audio with Inworld STT-1: synchronous for short clips, async jobs for long recordings, with speaker diarization, word timestamps, Voice Profile, and custom vocabulary. |
/inworld:add-streaming-stt [browser|server] |
Scaffolds a live mic → transcript streaming STT client. For captions, dictation, voice notes, meeting transcription, voice commands. Browser uses AudioWorklet; server uses SoX. |
/inworld:troubleshoot [error or problem] |
Diagnoses Inworld API errors and how-do-I questions using Inworld's official docs and support knowledge base (via search_docs), checks known sharp edges, reproduces with the MCP tools, and only then escalates to support — with a well-formed report. |
The plugin's MCP server exposes 10 tools that Claude can call directly:
| Tool | What it does |
|---|---|
list_voices |
Browse Inworld voices. Filters: language, description_match (keyword search of the voice description, e.g. ['warm'] — the reliable filter), tags (server-side tags, often empty today), custom_only. |
synthesize_speech |
TTS-2 by default; inworld-tts-2-flash for lowest latency. Inline steering, delivery_mode, WORD or CHARACTER timestamps. |
transcribe_audio |
Batch STT (file → text) with Inworld STT-1 (model_id: inworld/inworld-stt-1). Optional Voice Profile (age, emotion, pitch, vocal style, accent). |
clone_voice |
Clone a voice from one or more audio samples. Returns a new voiceId. |
design_voice |
Generate up to 3 preview voices from a text description. |
publish_voice |
Persist a designed voice to your library after picking the preview you like. |
chat_completion |
OpenAI-compatible call to any provider model (openai/gpt-4o-mini, anthropic/claude-3-5-sonnet, etc.) or to an Inworld Router (inworld/<name>). |
chat_completion_with_audio |
LLM + TTS in one call — chat reply returned as audio plus transcript. Faster than chat → synth because TTS pipelines while the LLM is still generating. |
list_routers |
List your configured Inworld Routers. |
search_docs |
AI search over Inworld's official knowledge: docs (docs portal + API reference), website (inworld.ai), resolutions (support knowledge base), ui-actions (Studio UI catalog). Public endpoint, no API key needed. |
Realtime, streaming TTS, and streaming STT are intentionally not MCP tools — bidirectional streaming audio belongs in your app, not in Claude's MCP transport. The corresponding skills write that code into your project.
Steering — inline in the text:
[whisper] meet me at the docks
[say with rising excitement] we did it!
[laugh] that was a close one
[speak slowly and with concern] are you sure you want to continue?
Delivery mode — overall consistency: STABLE, BALANCED (default), CREATIVE.
Pauses — SSML break tags:
Welcome. <break time="500ms"/> Let me help you with that.
Custom pronunciation — replace one word with its English IPA, wrapped in slashes:
Your interests are a perfect match for a honeymoon in /kriːt/.
Verbatim — spell out codes, IDs, and booking references character by character:
Your booking reference is <verbatim>AHAA7771Z</verbatim>.
Long text — 2,000 characters per synthesize call. For longer copy, split at sentence boundaries, or use async synthesis (preview, up to 100,000 characters as one background job). For large libraries of separate lines, batch synthesis (preview) costs 20% less per character.
OpenAI compatibility — apps already using OpenAI's text-to-speech can point the OpenAI SDK at https://api.inworld.ai/v1 and pick an Inworld model and voice.
- TTS-2 — 200+ languages and locales, natural-language steering (
[whisper],[say with rising excitement],[laugh]),delivery_mode, word & character timestamps - TTS-2 Flash — Inworld's fastest, lowest-cost model: ~20 ms time to first audio, same languages as TTS-2
- Streaming TTS — audio starts playing while it's still being generated, via chunked HTTP or WebSocket (~100 ms server-side time to first audio with TTS-2)
- STT — Inworld STT-1 in 30 languages, with optional Voice Profile (age, emotion, pitch, vocal style, accent) and custom vocabulary for names and jargon
- Long-recording transcription — async jobs for meetings, interviews, and podcasts (up to 512 MB), with speaker diarization ("who said what") and word timestamps
- Streaming STT — live mic → transcript over WebSocket, interim + final results, tunable or manual turn detection
- Realtime voice agents — bidirectional speech-to-speech over WebSocket / WebRTC / Twilio, with optional Router-backed LLM
- Instant voice cloning — a custom voice from a short sample in seconds; use 15–30 seconds of clean audio for the best similarity
- Professional voice cloning (beta) — a fine-tuned clone that's more similar and stable, trained on 10+ minutes of audio; created in the Inworld Portal or with the PVC API
- Voice design — generate voices from a text description (no audio sample required)
- LLM Router — OpenAI-compatible chat completions with traffic-split routing, fallback, and A/B testing
- LLM + TTS combined — one call: chat completion that returns audio directly
| Model | Languages | Latency | Best for |
|---|---|---|---|
inworld-tts-2 (default) |
200+ | ~100 ms TTFB | Quality, steering, multilingual |
inworld-tts-2-flash |
200+ (same as TTS-2) | ~20 ms TTFB | Lowest latency and cost; no steering instructions (non-verbal tags like [laugh] still work) |
STT (batch and streaming) uses inworld/inworld-stt-1: 30 languages, Voice Profile, and configurable turn-taking for streaming.
The MCP server is vendored from @inworld/cli (its in-tree packages/cli/src/mcp/ server, built as the inworld-mcp bin) — this repo has no server source of its own. The plugin pins the exact 10-tool surface via INWORLD_MCP_TOOLSET=runtime in .mcp.json, and the tool names + input schemas are locked by the CLI's tests/mcp/runtime-tools.test.ts contract test.
To change server behavior, change it in the CLI repo, then re-vendor:
cd inworld-tts-mcp
./scripts/vendor.sh # pinned npm version (see CLI_VERSION in the script)
./scripts/vendor.sh 1.3.1 # specific npm version
./scripts/vendor.sh ~/code/inworld-cli # local CLI checkout (must be built)
The script copies the artifact into build/index.js and runs scripts/smoke.mjs, which boots it standalone and asserts the 10-tool contract. The vendored build/index.js is committed so end-users don't need a build step.
The MCP server defaults to the production API (https://api.inworld.ai). To point it at a different environment (internal dev/staging), set INWORLD_API_BASE in the environment Claude Code is launched from:
export INWORLD_API_BASE="https://your-dev-endpoint.example.com"Trailing slashes are stripped. When unset or empty, production is used.
MIT