For your first 14 days you can skip this page entirely: press Start on the app's first screen and corrections run through RefineLoop's own connection: no key, no account, nothing to create. When you are ready for your own connection (free, and yours forever), the app walks you through it: Settings (Ctrl+Alt+O) → Get connected free. This page is the same guidance in long form, for reading ahead or helping somebody else.
The fastest way to have RefineLoop working today. Groq's free tier has no credits system and no card, just rate limits.
Refine Loopgsk_.groq, and paste the key.openai/gpt-oss-120b, click Test connection, then Save all.One key runs everything. Groq also serves speech-to-text (Whisper) and read-aloud (Orpheus voices) on the same free key, so dictation, voice messages, and natural read-aloud work immediately, with no Python and no extra deployments. Set “Speech-to-text runs” to Same as AI model in section 2 if you would rather not install Python. Free limits, measured on a real key: about 1,000 corrections, 2,000 dictations, and 8 hours of audio a day (far more than one person uses), but read-aloud is about 100 plays a day. When those run out, RefineLoop switches to the Windows voice by itself, so the sound never stops, it just gets plainer until the next day. Groq runs open-source models (Llama, Qwen, GPT-OSS), which handle English corrections well; a paid Gemini/OpenAI/Claude key is still noticeably better for Vietnamese explanations. Switching later is one dropdown.
Also free for the Flash models we use, though Google has been tightening this. If AI Studio asks you to set up billing, use Groq above instead; nothing in RefineLoop depends on which one you pick.
AIza… and is shown only once, so paste it somewhere safe.gemini, and paste the key into API key.gemini-2.5-flash, click Test connection, then Save all. That is it. Try Ctrl+Alt+B on a sentence.Be aware: on Google's free tier, your prompts may be used to improve their products. That is fine for practising English and everyday chat. For confidential work text, use a paid key or Ollama; the app treats every provider the same.
No key, no account, no internet needed after the download. RefineLoop sets it up for you; there is no terminal in this path. Speed is the catch: this needs a dedicated graphics card. On a laptop with built-in graphics a word lookup takes around 20 seconds, where a free cloud key answers in one or two.
ollama.llama3.1 (about 4.7 GB) is the recommended one, gemma3:4b (about 3.3 GB) suits smaller laptops.Local models are weaker than cloud ones, especially for Vietnamese and for the “why” explanations. Many people start on Gemini and switch to Ollama later for sensitive work.
The best quality, and the right answer for company data. Typical personal use costs well under a dollar a month.
openai, model gpt-4o-mini.claude, model claude-sonnet-5.azure, then copy the endpoint URL, KEY 1, and your deployment name from the Azure portal. The deployment name is what you typed when you deployed the model, not the model name.tts model and put its deployment name in section 3 for natural read-aloud voices, and a whisper model if you prefer cloud speech-to-text.Other OpenAI-compatible services work too: pick provider custom and paste their base URL. Groq (https://api.groq.com/openai/v1) and OpenRouter (https://openrouter.ai/api/v1) both have free tiers.
If you already run a local gateway that speaks the OpenAI API and routes to providers you chose, RefineLoop can use it like any other endpoint. This path is for people comfortable running a server. Nobody needs it, and the paths above stay simpler.
custom, and enter your gateway's address as the Endpoint: for OmniRoute that is http://localhost:20128/v1, with an API key from its dashboard and auto as the model.Two honest limits. RefineLoop treats a localhost endpoint as a local model, so it uses the shorter prompts written for laptop-class models even when your gateway routes to a big cloud one. Quality can be a shade below connecting to that provider directly. And privacy is what your gateway does, not what the address looks like: localhost here is a hop, and your text continues to whichever provider the gateway picks.
Dictation, voice messages, instant replay, and shadowing work out of the box.
Connection refused (localhost:11434) | Provider is set to ollama but Ollama is not running. Start Ollama, or switch the provider to gemini. |
DeploymentNotFound | Azure: the Deployment field must contain your deployment name from the Azure portal, not the model name. |
401 / Unauthorized | The API key is wrong, expired, or has a stray space. Paste it again, and use Show keys to check it. |
Nothing happens on a hotkey | Another app owns that shortcut. The tray menu shows a red warning; click it to rebind. |
Read-aloud sounds robotic | You are hearing the Windows fallback voice. Press Test voice in Settings section 3 - it says exactly why. On Groq, accept the voice model terms in their playground; on Azure, add a tts deployment. |
Speech features do nothing | Python is missing. Press a speech hotkey and choose Set it up to let the app install it; or see logs\whisper-server.log. |
Still stuck? Email support@refineloop.app with what you see on screen; a screenshot is perfect.