Use any other model

Path 1: Add a LiteLLM proxy

A LiteLLM proxy gives you one OpenAI-compatible endpoint in front of providers you already hold keys for: OpenAI, Azure, AWS Polly, Vertex AI/Gemini, ElevenLabs, or MiniMax. You host the proxy, not the models, so it runs fine on any machine.

1. Install it

You need Python 3.10 or newer. Then:

pip install 'litellm[proxy]'

2. Quick start with one backend

Export the backend's key and start the proxy with the model you want to route to. It listens on port 4000:

export OPENAI_API_KEY=sk-...
litellm --model openai/tts-1
# RUNNING on http://0.0.0.0:4000

3. Or route to several providers

Put each backend in a config.yaml. The model_name aliases you choose here are what you'll type into the extension's Models field:

model_list:
  - model_name: tts
    litellm_params:
      model: openai/tts-1
      api_key: os.environ/OPENAI_API_KEY
  - model_name: eleven
    litellm_params:
      model: elevenlabs/eleven_multilingual_v2
      api_key: os.environ/ELEVENLABS_API_KEY

# optional: require a key from clients (paste it into the
# extension's API key field)
general_settings:
  master_key: sk-your-proxy-key

litellm --config config.yaml

MiniMax, Vertex AI/Gemini, Azure, and Polly entries follow the same pattern; the exact model strings are in the LiteLLM TTS docs and its per-provider pages.

4. Point the extension at it

Server URL http://localhost:4000/v1; API key only if you set a master key; Models field tts, eleven (your aliases). For voices, OpenAI names like alloy work everywhere, and LiteLLM maps them per backend (for ElevenLabs, raw voice IDs also pass through unchanged).

Path 2: Host an open model (optional)

Skip this if the proxy path covers you. Hosting an engine yourself keeps every request on your machine and costs nothing per character. Common choices:

• LocalAI: a full local OpenAI replacement that loads multiple TTS backends, such as Kokoro, Piper, and Coqui
• Speaches: a lighter server that runs Kokoro and Piper voices behind one endpoint
• Kokoro-FastAPI: the quickest start if the Kokoro model is all you need
• openedai-speech: a small server that serves open TTS models on this API

For Kokoro-FastAPI, one Docker command starts a server on port 8880:

docker run -d --name kokoro -p 8880:8880 ghcr.io/remsky/kokoro-fastapi-cpu

The other engines document their own installs: see the LocalAI or Speaches getting-started pages. Most run as a Docker container or a Python package and listen on a local port.

Connect the extension

1. Click the extension icon in your browser toolbar
2. Open Settings → OpenAI-compatible
3. Paste the Server URL: http://localhost:4000/v1 for LiteLLM, http://localhost:8880/v1 for Kokoro-FastAPI, or your engine's own port
4. Enter an API key only if the server requires one (local servers usually don't)
5. Optionally list voice names separated by commas, and one or more models, also separated by commas (the default is tts-1, which most servers accept)
6. Click Save & test

The extension first asks the server for its voice list, so you can usually leave the voices field empty. If the server doesn't report voices, it falls back to OpenAI's nine standard voice names, and anything you type in the field takes priority. Each model you list becomes its own row per voice in the voice picker, so one server can expose several models at once and you can switch between them per read. Once the test passes, highlight text on any page, right-click, and choose Read aloud, or press Ctrl/Cmd+Shift+S.

Costs

The two paths bill differently. Through a LiteLLM proxy, each request is billed by the backend it routes to at that provider's own rates, and your text goes to that provider; LiteLLM itself is free software. A self-hosted engine is free to run once you have the hardware, with no per-character charge and nothing leaving your machine. For provider rates, see our provider comparison.

Remove access when you're done

1. Stop the server process (Ctrl+C) or delete its container.
2. In the extension, open Settings → OpenAI-compatible and clear the saved fields (or turn the provider off).