LLM Provider and Model

Switch providers or models in Settings → Account, with no restart. Your conversation history carries over.

• Anthropic Claude Sonnet — the default; best for complex multi-track arrangements and long planning chains.
• Anthropic Claude Haiku — fastest and cheapest; good for simple edits.
• OpenRouter — access to 50+ models including open-weight options at lower cost.
• OpenAI — reliable tool use; the default is GPT-6 Luna.
• Google Gemini and Groq — hosted alternatives, each with its own key.
• Ollama — a local model on your own machine; no key, and not even the chat leaves the computer.
• ChatGPT plan (a preview) — no key: the assistant runs on the usage included in your ChatGPT Plus or Pro plan. See Using your ChatGPT plan.

Your keys, and the ChatGPT sign-in, are kept together in one entry of your OS keychain, so macOS asks for your login password at most once per launch. Your provider, model, base URL and reasoning effort are not secret: they are saved in a small settings file in the app's config folder, which never holds a key. When you update from an earlier version, your saved keys move into the single entry on the first launch, which may ask once for each old entry.

With Anthropic selected, Settings → Account also has a Reasoning effort setting: Default, Low, Medium, High, Extra high or Max. Higher effort makes Claude think longer, which is slower and uses more tokens; Default leaves it to the model. The ChatGPT plan has its own, described in Using your ChatGPT plan; other providers have none.

At the foot of Settings → Account, one line names the ONNX Runtime library the app found, for example ONNX Runtime: 1.23.2 (bundled), with the file and the places it looked below it. It says not installed in this build when there is none, or why the one it found cannot be used, such as a macOS older than 13.4 (only the library needs it; the app runs on macOS 11 and later). An installed app says bundled; a build from source says not installed in this build. It is a status line: it installs nothing and does not mean any on-device audio model is available. No model comes with the runtime, and stem separation still returns an error.

The Speech model row below it offers Whisper base (about 139 MB, from Hugging Face). Download fetches its three files from huggingface.co, with a progress bar and a Cancel button, checks each against its SHA-256 hash, and keeps them in the models folder of the app’s local data folder (app.edytlab.desktop; on Windows under %LOCALAPPDATA%, so it does not roam). Cancel keeps what has arrived and Download carries on from there; a failure shows why, with Retry. When the row says Ready it offers Remove, which asks first and deletes only that model’s files. Nothing downloads unless you press Download: not at launch, not when Settings opens, not when the assistant asks. The requests carry your IP address, the file name, and a User-Agent of the form edytlab/<version>, so Hugging Face learns that edytlab is asking and which version, and nothing about your audio (see the privacy page).

The runtime’s licence and notices are installed with it, in licenses/onnxruntime/: inside edytlab.app/Contents/Resources on macOS, in /usr/lib/edytlab on Linux, and in the install folder on Windows, which also has the Visual C++ runtime the library needs and its notice in licenses/vc-runtime/.

The speech model has no notice in the install folder, since it is not installed with the app. Whisper base is published by OpenAI under the Apache-2.0 licence (the Whisper code is MIT); the files are an ONNX export by the onnx-community on Hugging Face, pinned to one commit. The model cards are openai/whisper-base and onnx-community/whisper-base_timestamped, the second of which states no licence of its own. edytlab fetches the files when you click Download and does not redistribute them.