Voice & model providers · Pinecall Docs

The speech-to-text, model and voice vendors an agent can run on, and bringing your own keys.

A call runs on three vendors: one that turns the caller's speech into text, a model that decides what to say, and a voice that says it. Pinecall runs any vendor LiveKit has a plugin for, and you pick them per agent, in the sandbox and in production, without a deploy.

What is a provider?#

Speech-to-textThe ears: stt. Turns the caller's audio into text as they speak. ModelThe llm. Reads the prompt and the conversation, and answers or calls a tool. VoiceThe tts. Says the answer, in the voice you chose.

A provider is a LiveKit Agents plugin installed on the gateway's workers. Nothing in Pinecall's code names a vendor: every installed plugin that offers an STT, an LLM or a TTS is a provider. Besides them, livekit is LiveKit Inference, which fronts many models under one name.

What's installed#

On Pinecall Cloud the workers carry these plugins, besides livekit:


`anthropic`	`assemblyai`	`asyncai`	`baseten`	`bland`	`cambai`
`cartesia`	`cerebras`	`clova`	`deepgram`	`elevenlabs`	`fal`
`fireworksai`	`fishaudio`	`gladia`	`gnani`	`gradium`	`groq`
`hume`	`inworld`	`lmnt`	`meta`	`minimax`	`mistralai`
`murf`	`neuphonic`	`nvidia`	`openai`	`palabra`	`perplexity`
`resemble`	`respeecher`	`rime`	`rtzr`	`sarvam`	`simplismart`
`slng`	`smallestai`	`soniox`	`speechify`	`spitch`	`upliftai`
`vakyam`	`xai`

AWS, Azure, Google and Speechmatics are an optional install that a self-hosted runtime can add (Self-hosting).

The exact list, what each vendor does, and whether it would run a call for your org now, is one command:

Terminal

$ pinecall providers --does stt

Its last line names the vendors your org runs on when an agent sets none.

Choosing them for an agent#

The vendors are settings of the agent, in the sandbox and in production, or declared in the class, which then wins:

Terminal

$ pinecall agent set --stt deepgram --llm anthropic/claude-haiku-5-5 --tts cartesia --prod
$ pinecall voices --tts cartesia --language es

--llm takes vendor/model, a vendor alone (its own default model), or a model alone (on the vendor in use). A vendor that is not installed, or does not offer that stage, is refused when you set it. Agent settings has every knob, the plugin's own class and options among them; those run only on your own key for the vendor.

Which are priced#

Every call carries its cost, priced per vendor and model that actually answered. Pinecall ships prices for these:

vendor	priced
`anthropic`	Claude models, by tokens
`openai`	GPT, realtime, transcription and voice models
`deepgram`	Nova and Flux speech-to-text, Aura voices
`elevenlabs`	its voices, and Scribe speech-to-text
`cartesia`	Sonic voices, Ink speech-to-text
`soniox`	speech-to-text
`perplexity`	Sonar models
`twilio`	phone minutes, for calls through Twilio

A vendor with no price still runs. Its usage is recorded, and the call's cost lists it as unpriced.

Bringing your own keys#

With no key of your own, a vendor runs on Pinecall's key, for the vendors and models Pinecall lends your org. Bring your own key and every call of the org runs that vendor on it, any model of it:

Terminal

$ pinecall providers add elevenlabs     # the key on stdin, never as a flag
$ pinecall providers list               # the vendors you brought, names only
$ pinecall providers rm elevenlabs      # back to Pinecall's key

No command or screen ever shows a key back. Limits & provider keys explains what ready, no key and no plugin mean, and who may add keys.

When a default vendor fails during a call, Pinecall can switch that stage to a fallback the gateway's operator configured, and the call's log records a vendor.switched entry. An agent that names its own vendor runs that vendor alone, with no fallback.

Next steps#

Agent settingsThe voice, the model and the ears of one agent. Limits & provider keysWhat Pinecall lends, and your own keys. Call metricsLatency per stage and what each call cost. Provider keys APIThe catalog and the keys, at the gateway.