Inworld AI
Inworld AI provides native APIs for speech synthesis, speech recognition, voice management, model discovery, embeddings, knowledge, and router management.
Verdict
Common use cases
- Prototype voice UI flows with realistic speech
- Transcribe customer call recordings for analysis
- Generate embeddings for semantic document search
- Preview multilingual voices before production
- Convert meeting notes into audio summaries
Integration
- Vendor
- Inworld AI
- Category
- other
- Auth
- API_KEY
- Tools
- 6
- Composio slug
inworld_ai
Tools
- Create Embeddings
Generate one complete embedding vector per input text with an Inworld-hosted embedding model. Batches can produce large responses.
- Get Voice Preview
Generate the provider-selected, unmetered MP3 preview for a voice and return it as a downloadable file.
- List Models
List Inworld-supported models and the pricing, modality, context, and output limits needed to choose a model. The endpoint returns the complete model catalog and can produce a moderately large response.
- List Voices
Return one page of voices available for preview or speech synthesis, with optional filtering, language selection, and sorting. The filter and languages parameters are mutually exclusive.
- Synthesize Speech
Synthesize short text with an Inworld voice and return the complete audio as a downloadable file. Long text can exceed tool payload limits; prefer short text and compressed MP3 output.
- Transcribe Audio
Synchronously transcribe a short audio file and return the final transcript, timestamps, and usage. Prefer compressed input to limit base64 payload growth.
Setup
Setup guide
- 11. In Switchy, open your workspace settings and navigate to the Integrations tab. 2. Click 'Add MCP Integration' and select Inworld AI from the catalog. 3. Go to the Inworld AI developer console, create a new API key, and copy it to your clipboard. 4. Paste the key into Switchy's configuration dialog and click 'Connect'. 5. Switchy will verify the key by listing available models; if successful, you'll see a green checkmark. 6. Open any Project, type '@Inworld' in the chat, and ask it to list voices or synthesize a short phrase. 7. If the response includes a downloadable MP3 or transcript, the connection is working. 8. For transcription, upload a short audio file to the Project's file panel first, then reference it by name in your prompt. 9. Keep text inputs under a few hundred words and audio files under five minutes to avoid payload errors.
What teammates see: by default, memories from Inworld AI are scoped to the Project (PROJECT visibility) - you can mark any memory PRIVATE or share it ORG-wide.
Works well with
Top models
Compatibility data appears once enough Projects have used this MCP together with a given model.
How Switchy teams use it
Starter prompts
Preview Voice Options
@Inworld list voices available in English, filtered for female speakers, and generate a preview MP3 for the top three results.
Synthesize Product Demo Script
@Inworld synthesize the following script using a professional male voice: 'Welcome to the dashboard. Here you can track your team's progress in real time.' Return the audio as an MP3.
Transcribe Interview Recording
@Inworld transcribe the audio file 'interview_2024.mp3' and return the full transcript with timestamps for each segment.
Generate Text Embeddings
@Inworld generate embeddings for these three product descriptions: 'wireless headphones', 'noise-canceling earbuds', 'studio monitor speakers'. Return the vectors.
Compare Model Options
@Inworld list all supported models and show me the pricing, modality, and context window for each one.
Example outputs
Illustrative - representative of the model's voice and quality, not literal recordings.
@inworld list available voices filtered by 'female' and sort by name, then preview the top result
Retrieved 47 female voices from the Inworld catalog, sorted alphabetically. Top result: 'Aria (en-US)' — a warm, conversational voice suitable for narration and customer service contexts. Generated a 3-second MP3 preview demonstrating tone and cadence. The preview file (24KB) is ready for download. Note that this is the provider's unmetered sample; actual synthesis requests will consume quota based on character count and selected voice tier.
This example shows voice discovery and preview generation — useful when selecting voices for a project before committing to synthesis. The filter and sort parameters help narrow large catalogs quickly. Be aware that the 'languages' and 'filter' parameters are mutually exclusive; you cannot combine them in a single request.
@inworld synthesize speech from this script: 'Welcome to the onboarding session. Let's walk through your first workflow.' Use a neutral English voice and return MP3 format.
Synthesized 14 words using voice 'Jordan (en-US, neutral)' and returned a 6-second MP3 file (48KB, 24kHz sample rate). Audio is ready for download. This request consumed approximately 70 characters from your synthesis quota. For longer scripts exceeding ~500 characters, consider breaking text into smaller chunks to avoid payload size limits in the MCP tool response.
Demonstrates text-to-speech synthesis with a practical onboarding script. Inworld's synthesis tool works best with short inputs (under 500 characters) due to base64 encoding overhead in MCP responses. For production workflows with longer content, plan to batch requests or use Inworld's streaming API directly outside the MCP context.
@inworld transcribe this audio file and create embeddings for the resulting transcript to enable semantic search across our voice memo library
Transcribed a 22-second audio file (MP3, 16kHz) and extracted the text: 'Discussed Q3 roadmap priorities including API rate limit improvements and new webhook events for real-time notifications.' Transcription included word-level timestamps for alignment. Then generated a 1536-dimension embedding vector for the transcript text, suitable for cosine similarity search in a vector database. Total processing time: ~4 seconds. Compressed input formats (MP3, Opus) are strongly recommended to keep base64 payloads manageable.
This example chains two capabilities — transcription and embedding generation — to build a searchable voice memo index. The workflow is practical for small audio files but note that transcription is synchronous and may time out on files longer than 60 seconds. Embedding responses can be large (1536 floats); ensure your downstream storage can handle the vector dimensions returned by Inworld's model.
Use-case deep-dives
When Inworld AI speeds up character voice iteration
A 6-person indie game studio needs to test 20 NPC voice profiles before recording final VO. The team uses Inworld AI's voice preview and synthesis tools to generate MP3 samples in standup, then shares them in Slack threads for director feedback. This MCP is the right call when you're iterating on character tone and need quick turnarounds—List Voices filters by language and style, Synthesize Speech handles short barks and one-liners, and Get Voice Preview lets you audition provider samples without burning synthesis credits. The threshold: if your script lines exceed a few sentences, the tool payload limit becomes a problem and you'll need to batch outside Switchy. For early-stage character work where speed beats polish, this MCP closes the loop between writing and hearing in under a minute.
Why Inworld AI's transcription fits small support teams
A 3-person support team at a SaaS startup records Zoom calls with enterprise customers and needs transcripts for ticket context. The Transcribe Audio tool handles compressed MP3 uploads and returns timestamped text in one synchronous call, which the team pastes into Linear issues during post-call triage. This MCP works when your audio files are under 10 minutes and already compressed—longer calls hit base64 payload limits and you'll need a dedicated transcription service. The API key auth means anyone on the team can run transcriptions without OAuth friction, and the synchronous response keeps the workflow inside a single Switchy thread. If your call volume is under 50 per week and you're already compressing audio for storage, Inworld AI's transcription is faster than spinning up a Whisper instance or paying for a per-seat transcription SaaS.
When Inworld AI embeddings replace a vector DB setup
A 10-person product team maintains 200 Notion pages of product specs and wants semantic search without standing up Pinecone or Weaviate. The Create Embeddings tool generates vectors for each doc chunk, which the team stores in a simple Postgres table with pgvector. This MCP is the right call when you're building a one-off search prototype or your doc corpus is under 10k chunks—List Models shows you the embedding dimensions and context limits, and the batch endpoint keeps token costs predictable. The trade-off: if your corpus grows past 50k chunks or you need sub-second query latency, you'll outgrow this approach and need a purpose-built vector store. For teams that want semantic search this quarter without a dedicated infra sprint, Inworld AI's embeddings get you from zero to working search in a day.
Frequently asked
What does the Inworld AI MCP do in Switchy?
It connects your team to Inworld's voice, transcription, and embedding models. You can synthesize speech from text, transcribe audio files, generate embeddings for semantic search, and preview voices — all from Switchy's chat interface. Think of it as a speech-and-language toolkit you call through prompts instead of writing API code.
Do I need an Inworld AI account to use this MCP?
Yes. You'll need an Inworld AI account and an API key. The MCP uses API_KEY authentication, so whoever connects it in Switchy must paste their key into the connection settings. Inworld bills usage separately; the MCP just routes requests through your existing account.
Can this MCP generate long-form audio or handle large files?
No. The Synthesize Speech and Transcribe Audio tools are designed for short inputs to avoid payload size limits. If you need to process long audio or generate multi-minute speech, you'll hit encoding constraints. For bulk work, use Inworld's API directly or split jobs into smaller chunks.
How is this different from calling Inworld's API myself?
The MCP wraps Inworld's endpoints so your team can trigger them conversationally in Switchy without writing code. You lose fine-grained control over parameters and can't stream responses, but you gain speed for ad-hoc tasks like testing voices or transcribing a quick file. For production pipelines, the API is still the better choice.
Who on my team should connect the Inworld MCP?
Whoever holds the Inworld API key and understands your usage budget. The MCP doesn't enforce role-based access, so anyone in the Switchy workspace can call the tools once connected. If you're worried about cost control, connect it under a dedicated service account and monitor usage in Inworld's dashboard.