AI providers
Configure OpenAI-compatible transcription and summarization providers.
OpenPlaud uses the OpenAI SDK with a configurable baseURL. Any provider
that exposes an OpenAI-compatible API works: OpenAI itself, Groq,
OpenRouter, Together, Azure OpenAI, LM Studio, Ollama, and anything else
that speaks the same wire format.
You can configure multiple providers and pick which one runs transcription and which runs summarization on a per-recording basis.
Adding a provider
Settings → AI providers → Add provider opens a dialog with a preset
list. Each preset pre-fills baseURL, default model, and a hint about
which API surface the transcription worker should use:
- Whisper-style (
/v1/audio/transcriptions, multipart upload) — the classic OpenAI Whisper surface. Used by OpenAI, Groq Whisper, Together Whisper, LM Studio, Ollama. - Chat-style (
/v1/chat/completionswith aninput_audiocontent part) — for providers that expose audio-input LLMs instead of a dedicated transcription endpoint. Used by OpenRouter routing Gemini / GPT-audio / Voxtral.
Pick a preset, paste your API key, choose a default model, save. Repeat for additional providers.
Preset reference
| Preset | Base URL | Notes |
|---|---|---|
| OpenAI | (blank, defaults to api.openai.com) | Production quality; whisper-1, gpt-4o-transcribe |
| Groq | https://api.groq.com/openai/v1 | Free Whisper API, very fast |
| OpenRouter | https://openrouter.ai/api/v1 | Chat-style transcription; many summarization models in one key |
| Together AI | https://api.together.xyz/v1 | Cost-effective, broad model selection |
| Azure | https://<resource>.openai.azure.com/... | Enterprise compliance, deployment-scoped URLs |
| LM Studio | http://localhost:1234/v1 | Local, 100% offline |
| Ollama | http://localhost:11434/v1 | Local, easy model management |
| Custom | (freeform) | Anything else OpenAI-compatible |
LM Studio and Ollama presets default to http://localhost. From inside
a Docker container, localhost is the container itself, not the host
machine — use http://host.docker.internal:<port> on Mac/Windows, or
the Docker bridge IP on Linux, to reach an AI server running on the
host. If you run Ollama in the same Docker network as OpenPlaud, use the
service hostname (http://ollama:11434/v1).
Choosing models
The preset list ships with a curated set of known transcription model ids per provider. When the API exposes structured capability metadata (today only OpenRouter), the Add/Edit dialog can fetch the live list of audio-capable models for you. For everyone else, the curated list drives the dropdown and a Custom… option lets you type an id by hand for new releases that ship before our next version bump.
Summarization is more forgiving — any chat model that follows the
OpenAI chat/completions shape works. Pick something fast and cheap
(e.g. gpt-4o-mini, llama-3.1-70b-versatile) unless you have a
specific reason to use a larger model.
What gets sent to the provider
Plaintext audio (for transcription) and plaintext transcript text (for summarization). The trust boundary is the provider you picked. OpenPlaud's at-rest encryption protects the database, not the network hop to your AI provider — see Encryption at rest for the explicit list of what that layer does and does not cover.
If you need to minimise that trust boundary, run a local provider (LM Studio or Ollama) on the same machine as a self-host instance. Both transcription and summarization stay on your hardware.
Switching the active provider
On any recording, the workstation panel exposes a transcription provider picker and a summary provider picker. Changes apply immediately; you can re-run a transcription with a different provider if the first result was poor.
Last updated on