Use case
A local-first AI audio editor
edytlab does not upload your audio. The editing runs on your computer, the AI only sequences the tools, and with Ollama even the chat can stay at home.
What stays on your machine, and what leaves
| Item | Where it goes |
|---|---|
| Your audio files, and the edits made to them (EQ, fades, time-stretch, export) | Nowhere. Decoded, processed and rendered on your computer. |
| Your chat messages and the agent's tool results | The AI provider you chose. Never your raw audio. A transcript is one of these results: the words the speech model found (a bounded amount for each call) reach the provider as text, and with a local Ollama model they stay on your computer too. |
| API keys | Your operating system's credential store, sent only to the provider that owns the key. The ChatGPT plan has no key. |
| Signing in with ChatGPT (optional) | Only if you click Continue with ChatGPT. Your browser goes to auth.openai.com, which edytlab also uses to exchange and renew the sign-in and to revoke it when you sign out. The assistant then talks to api.openai.com with your ChatGPT access token instead of a key: the same text as for any provider, never your raw audio. The refresh token stays in your operating system's credential store, and a small registration file (host ID, client ID, your email as a label) stays in the app's config folder. Signing out in edytlab does not disconnect it in ChatGPT; the privacy page says how. |
| The speech model (about 139 MB) | Downloaded from Hugging Face (huggingface.co) only when you click Download in Settings, never automatically. Hugging Face sees your IP address, the file you asked for and that the request comes from edytlab (and which version), and nothing about your audio. Transcribing with it runs on your computer and makes no request of its own. |
| The speaker model (about 33 MB) | Downloaded from Hugging Face (huggingface.co) only when you click its Download in Settings, never automatically. Hugging Face sees your IP address, the files you asked for and that the request comes from edytlab (and which version), and nothing about your audio. Finding who speaks when with it runs on your computer and makes no request of its own. |
| The stem model (about 171 MB) | Downloaded from Hugging Face (huggingface.co) only when you click its Download in Settings, never automatically. Hugging Face sees your IP address, the files you asked for and that the request comes from edytlab (and which version), and nothing about your audio. Splitting a track into stems with it runs on your computer and makes no request of its own; the stems are files in your project, and the assistant is told where they are, like any tool result. |
| Usage data from the app | None. The app sends no telemetry. |
| Visits to this website | Recorded by anonymous Vercel analytics (Web Analytics and Speed Insights), which the website loads and the app does not. See the privacy page. |
The privacy page has the details, and the source is public, so you can check the claims.
Why the AI does not need your audio
The model never sees or hears samples. It is given a description of the session (tempo, key, beat grid, loudness) and a list of deterministic audio tools, and chooses which to call and in what order. The tools do the signal processing on your machine, so the private part stays local and the part that needs a language model is only text.
Run it with Ollama
Ollama needs no account and no key. Install Ollama and pull a model that supports tool calling, then choose Ollama in edytlab's Settings and press Test, which sends a minimal request to check that the model can call tools. The getting started guide walks through it, and you can switch between Ollama and a hosted provider at any time, without reinstalling.
Be realistic about local models
- With Ollama the agent sends a smaller set of tools: the common editing tools, plus any you have named or the assistant has just proposed. The names of the rest go in the prompt, so the model can still call them. A big project's session summary is cut to fit too, and the model reads the rest a page at a time. Ollama trims an over-long prompt and the app shows nothing, so if replies start to forget earlier messages, raise the model's context length or reopen the project to start the conversation afresh. The Ollama post has the limits in full.
- On a CPU, a large request means a long wait for the first reply.
- “Plan first” mode, which shows the agent's steps for your approval before anything runs, is a useful guard with smaller models.
- Transcription and stem separation run on your computer's CPU, whichever model you chat with. Stem separation needs about 4 GB of free memory and takes roughly half the length of the audio.
Where next
See it work in the DJ blend demo or browse the tools reference. The builds are unsigned for now, so macOS and Windows will warn on first launch; the setup guide has the steps.