VoiceGateway uses a provider registry pattern that makes adding new providers straightforward. Each provider is a single Python file that extends `BaseProvider`. This guide walks through the full proc

Adding a Provider

VoiceGateway uses a provider registry pattern that makes adding new providers straightforward. Each provider is a single Python file that extends BaseProvider. This guide walks through the full process.

Prerequisites

Development environment set up
Familiarity with the provider's API (SDK, auth, pricing)
Account with the provider for testing

10-step checklist

1. Create the provider file

Create src/voicegateway/providers/<name>_provider.py. Use an existing provider as a template (e.g., deepgram_provider.py for STT, cartesia_provider.py for TTS).

Python

"""<Provider Name> provider implementation."""

from __future__ import annotations

from typing import Any

from voicegateway.providers.base import BaseProvider


class <Name>Provider(BaseProvider):
    """<Provider Name> provider for <STT/LLM/TTS>."""

    def __init__(self, config: dict[str, Any]) -> None:
        self._api_key = config.get("api_key", "")
        # Initialize any SDK client here

    def create_stt(self, model: str, **kwargs: Any) -> Any:
        """Create an STT instance, or call self._unsupported('stt')."""
        self._unsupported("stt")

    def create_llm(self, model: str, **kwargs: Any) -> Any:
        """Create an LLM instance, or call self._unsupported('llm')."""
        self._unsupported("llm")

    def create_tts(self, model: str, voice: str | None = None, **kwargs: Any) -> Any:
        """Create a TTS instance, or call self._unsupported('tts')."""
        self._unsupported("tts")

    async def health_check(self) -> bool:
        """Check if the provider is reachable."""
        # Make a lightweight API call to verify connectivity
        return True

2. Implement the BaseProvider ABC

The BaseProvider abstract class in src/voicegateway/providers/base.py requires four methods:

Method	Purpose	Return type
`create_stt(model, **kwargs)`	Create an STT plugin instance	LiveKit STT plugin or `None`
`create_llm(model, **kwargs)`	Create an LLM plugin instance	LiveKit LLM plugin or `None`
`create_tts(model, voice, **kwargs)`	Create a TTS plugin instance	LiveKit TTS plugin or `None`
`health_check()`	Verify provider connectivity	`bool`

For modalities the provider does not support, call self._unsupported("modality_name") to raise a clear error.

Pricing is not a provider-level concern. LLM, STT, and TTS rates all resolve via voice-prices (the wrappers live at src/voicegateway/pricing/{llm,stt,tts}.py). To add pricing for a new model, see step 4.

3. Register the provider

Add your provider to the registry in src/voicegateway/core/registry.py:

Python

_PROVIDER_REGISTRY: dict[str, tuple[str, str]] = {
    # ... existing providers ...
    "<name>": ("voicegateway.providers.<name>_provider", "<Name>Provider"),
}

The registry uses lazy imports -- your provider module is only loaded when a user configures it. This means optional dependencies do not break the install.

4. Add pricing data

Pricing for every modality (LLM, STT, TTS) resolves through voice-prices, so there is no VoiceGateway-side catalog entry to add. Confirm the model id resolves with the matching wrapper:

Python

from voicegateway.inference.pricing import llm, stt, tts

llm.calculate_llm_cost("<name>/<model>", 1000, 500)   # LLM
stt.calculate_stt_cost("<name>/<model>", 60)          # STT (audio seconds)
tts.calculate_tts_cost("<name>/<model>", 1000)        # TTS (characters)

If a call returns None, the model is not yet in voice-prices. Add it upstream in voice-prices (each entry carries prices_checked and pricing_source_url), publish a new voice-prices version, and bump the pin in pyproject.toml. Self-hosted models (local/*, ollama/*) price at $0 automatically and need no entry.

5. Add optional dependency

Add a new extra in pyproject.toml:

TOML

[project.optional-dependencies]
<name> = ["<sdk-package>>=1.0.0"]

# Update the cloud or local group as appropriate
cloud = ["voicegateway[deepgram,openai,...,<name>]"]

6. Add fake API key to test fixtures

In src/voicegateway/tests/conftest.py, add the key to the _test_env fixture:

Python

@pytest.fixture(autouse=True)
def _test_env(monkeypatch):
    for key in [
        # ... existing keys ...
        "<NAME>_API_KEY",
    ]:
        monkeypatch.setenv(key, "test-key-value")

7. Write tests

Create src/voicegateway/tests/test_<name>_provider.py:

Python

"""Tests for the <Name> provider."""

import pytest

from voicegateway.providers.<name>_provider import <Name>Provider


@pytest.fixture
def provider():
    return <Name>Provider({"api_key": "test-key"})


def test_create_stt(provider):
    # Test STT creation or verify it raises NotImplementedError
    ...


def test_create_llm(provider):
    ...


def test_create_tts(provider):
    ...


async def test_health_check(provider):
    # Mock the HTTP call
    ...


def test_pricing_resolves_stt(provider):
    """Pricing for a known STT model resolves to a positive Decimal via the catalog."""
    from voicegateway.pricing import catalog
    cost = catalog.calculate_cost("stt", "<name>/model-name", audio_seconds=60)
    assert cost is not None and cost > 0


# For an LLM provider, dispatch with token kwargs:
#
#     cost = catalog.calculate_cost(
#         "llm", "<name>/model-name", input_tokens=1000, output_tokens=500
#     )
#
# For a TTS provider, use character_count:
#
#     cost = catalog.calculate_cost(
#         "tts", "<name>/model-name", character_count=100
#     )

See the testing guide for mock patterns and fixture usage.

8. Update documentation

Add the provider to relevant documentation pages:

docs/guide/what-is-voicegateway.md -- provider list
docs/guide/installation.md -- extras table
docs/configuration/ -- config example
README.md -- provider count and list

9. Test the full flow

Shell

# Lint
ruff check src/voicegateway/providers/<name>_provider.py

# Type check
mypy -p voicegateway.providers.<name>_provider

# Run your tests
pytest src/voicegateway/tests/providers/test_<name>_provider.py -v

# Run the full suite to check for regressions
pytest

10. Open a PR

Create a PR with:

Title: feat(providers): add <Provider Name> support
Description: what modalities are supported, link to provider docs, pricing source
Checklist: all items from the contributing guide

Example: anatomy of an existing provider

Looking at the registry, VoiceGateway ships with these 11 providers:

Provider	Module	Class	Modalities
openai	`openai_provider`	`OpenAIProvider`	STT, LLM, TTS
deepgram	`deepgram_provider`	`DeepgramProvider`	STT, TTS
anthropic	`anthropic_provider`	`AnthropicProvider`	LLM
groq	`groq_provider`	`GroqProvider`	STT, LLM
cartesia	`cartesia_provider`	`CartesiaProvider`	TTS
elevenlabs	`elevenlabs_provider`	`ElevenLabsProvider`	TTS
assemblyai	`assemblyai_provider`	`AssemblyAIProvider`	STT
ollama	`ollama_provider`	`OllamaProvider`	LLM
whisper	`whisper_provider`	`WhisperProvider`	STT
kokoro	`kokoro_provider`	`KokoroProvider`	TTS
piper	`piper_provider`	`PiperProvider`	TTS

Adding a Provider

On this page