ElevenLabs (Audio)

Speech-to-Text (Scribe) and Text-to-Speech through ClawRouter. ElevenLabs is a passthrough audio provider -- requests are forwarded byte-identical, with your managed API key injected automatically.

Version 1.0.18

Not a chat provider: ElevenLabs has no chat completions. Chat-style requests (/v1/chat/completions, /v1/messages, etc.) sent to an ElevenLabs provider fail with HTTP 400 -- format translation does not apply.

The ElevenLabs Preset

Setting	Value
Template	Quick Setup > ElevenLabs
Upstream URL	`https://api.elevenlabs.io/v1`
API Format	`elevenlabs`
API Key Mode	`Managed`
Get API Key	elevenlabs.io/app/settings/api-keys

Seeded models: scribe_v2, scribe_v1 (Speech-to-Text) and eleven_v3, eleven_multilingual_v2, eleven_flash_v2_5, eleven_turbo_v2_5 (Text-to-Speech). Use Fetch Models on the Models tab to pull the live TTS model list -- the Scribe STT models are always included. See Available Models below.

Authentication

ElevenLabs requires the xi-api-key: <key> header -- a Bearer or x-api-key credential is rejected upstream with HTTP 401 (invalid_authorization_header). ClawRouter handles this for you in Managed key mode:

Your client authenticates to ClawRouter with the normal proxy API key (Settings > Proxy API Key).
ClawRouter strips that credential and injects your managed ElevenLabs key as xi-api-key on the way out.

Key rotation, backoff, and connection testing work exactly as on chat providers.

Using the Proxy

Point your client (or curl) at the provider's Base URL and call the standard ElevenLabs v1 endpoints:

text

http://localhost:3030/proxy/{provider-id}/v1

Endpoint	Purpose
`POST /v1/text-to-speech/{voice_id}`	Text-to-Speech -- binary audio, piped byte-identical
`POST /v1/text-to-speech/{voice_id}/stream`	Streaming TTS -- zero-buffering audio pipe
`POST /v1/speech-to-text`	Speech-to-Text (multipart/form-data, JSON transcript)
`GET /v1/voices`	List available voices
`GET /v1/models`	List TTS models
`GET /v1/user`	Account info (also used by the connection test)

Text-to-Speech Example

bash

curl -X POST http://localhost:3030/proxy/{provider-id}/v1/text-to-speech/{voice_id} \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer cr_your_proxy_key" \
  -d '{
    "text": "Hello from ClawRouter!",
    "model_id": "eleven_multilingual_v2"
  }' \
  --output speech.mp3

Speech-to-Text Example

bash

curl -X POST http://localhost:3030/proxy/{provider-id}/v1/speech-to-text \
  -H "Authorization: Bearer cr_your_proxy_key" \
  -F "model_id=scribe_v2" \
  -F "file=@/path/to/audio.mp3"

Returns the JSON transcript from the Scribe model.

Available Models

Speech-to-Text (Scribe): use via POST /v1/speech-to-text (model_id field). Scribe models are not listed by the /v1/models endpoint, so ClawRouter always keeps them in the seeded list.

Model ID	Notes
`scribe_v2`	Latest Scribe STT
`scribe_v1`	Previous Scribe STT

Text-to-Speech: pass as model_id in the TTS request body.

Model ID	Notes
`eleven_v3`	Latest flagship TTS
`eleven_multilingual_v2`	Multilingual, high quality
`eleven_flash_v2_5`	Low latency
`eleven_turbo_v2_5`	Balanced latency / quality

Model lists change over time -- snapshot from August 2026. Models tab > Fetch Models pulls the live TTS list (GET /v1/models). Official catalog: elevenlabs.io/docs/models

Connection Testing

The Test button uses a zero-cost GET /v1/user probe -- no characters are billed. A definitively invalid key (401/403, or hard billing) is auto-disabled, exactly as on chat providers.

Limitations

No chat / no format translation. elevenlabs is a non-chat passthrough format. Chat requests fail with HTTP 400.
No Quota tab / usage probe. ElevenLabs has no supported usage endpoint.
No token or cost extraction. Audio responses carry no token usage, so these requests show no token counts or estimated cost in the logs. Binary audio bodies are never stored in logs -- a placeholder is recorded instead.
Realtime STT/TTS is not proxied. ElevenLabs realtime endpoints run over client-direct WebSocket (wss://api.elevenlabs.io/...) -- connect to ElevenLabs directly for those. (Client-side vs server-side streaming are commit strategies over the same realtime endpoint; both bypass the proxy.)