Connect a model

mi·do·na doesn't ship a model of its own. You bring one and connect it from the model connection panel in the top bar. Two ways to go: run a model locally, so nothing leaves your machine, or point mi·do·na at a cloud provider, where the model runs on someone else's hardware. Here's each service we support.

Open the connection panel

Click MODEL in the top-right of the app. Pick a service, fill in its details, and load the available models. Your connection details and any keys are stored in your browser only, never on our side.

Trouble connecting to a local model? If mi·do·na is served over HTTPS, the browser can block it from reaching http://localhost - most often in Safari. Try wrapping your runner in a proxy:

Reaching a local model over HTTPS

LM Studio

A model running locally on your own machine. Nothing leaves your computer. You'll need LM Studio installed with a model loaded and its server running.

ServiceChoose in the panel: LM Studio.
API endpointDefault: http://localhost:1234
Load modelsThen pick one: mi·do·na asks the server what's loaded and lists it. Select the model and you're connected.

New to running a model locally? Start with Set up your model runner. If an HTTPS page won't reach localhost (most often in Safari), you'll also need an HTTPS proxy.

Ollama

A model running locally on your own machine. Nothing leaves your computer. Ollama runs a server in the background as soon as it's installed - just pull a model and allow browser access.

ServiceChoose in the panel: Ollama, or Generic OpenAI.
API endpointDefault: http://localhost:11434
Load modelsThen pick one: mi·do·na lists what you've pulled. Select one and you're connected.

Need to install it? See Set up Ollama. Remember to set OLLAMA_ORIGINS so the browser can reach it.

llama.cpp

A model running locally on your own machine. Nothing leaves your computer. Its llama-server program serves an OpenAI-compatible API, with permissive CORS on by default.

ServiceChoose in the panel: llama.cpp, or Generic OpenAI.
API endpointDefault: http://localhost:8080
Load modelsThen pick one: mi·do·na lists the loaded model. Select it and you're connected.

Need to install it? See Set up llama.cpp.

koboldcpp

A model running locally on your own machine. Nothing leaves your computer. A single-file executable that serves an OpenAI-compatible API, with permissive CORS on by default.

ServiceChoose in the panel: koboldcpp, or Generic OpenAI.
API endpointDefault: http://localhost:5001
Load modelsThen pick one: mi·do·na lists the loaded model. Select it and you're connected.

Need to install it? See Set up koboldcpp.

Generic OpenAI

A catch-all for any server that speaks the OpenAI API. Point mi·do·na at a URL, add a key if the server wants one, and you get the basics. This is also how the other local runners connect - Ollama, llama.cpp, and koboldcpp each serve an OpenAI-compatible endpoint on their own port. Local or hosted depends entirely on the URL you give it.

ServiceChoose in the panel: Generic OpenAI.
API endpointAny OpenAI-compatible URL: The server, with no route on it - e.g. http://localhost:11434 for Ollama, http://localhost:8080 for llama.cpp, http://localhost:5001 for koboldcpp, or a hosted provider's URL.
API keyIf required: Local runners usually need none; a hosted endpoint will. Stored in your browser and sent only to the URL you set.
Load modelsThen pick one: mi·do·na asks the server what it offers and lists it. Expect the basics - exotic provider-specific features may not carry over.

OpenRouter

Cloud, not private. A hosted gateway to many models behind one API key. The model runs on OpenRouter's side, not yours, so this isn't a local or private option - whoever hosts the model can see what you send. You bring the key; requests go straight from your browser to OpenRouter, not through us.

ServiceChoose in the panel: OpenRouter.
API keyFrom your account: Create a key at openrouter.ai/keys and paste it in. It's stored in your browser and sent straight to OpenRouter over HTTPS when you make a request. We never see it.
Load modelsThen pick one: The full catalog loads. Pick a model; note that hosted models bill against your OpenRouter credit.

AI Horde

Cloud, not private. A free, community-run pool of volunteered GPUs. The model runs on a stranger's machine, so treat it as public - fine for casual use, not for anything sensitive. No cost, but you wait in a queue and speed depends on who's online. An anonymous key works; a registered one gets priority.

ServiceChoose in the panel: AI Horde.
API keyOptional: Leave blank to use the shared anonymous key (0000000000), or register at aihorde.net for a key with priority.
Load modelsThen pick one: Available models depend on what volunteers are hosting right now, so the list changes over time.

Whoever runs the model you connect can see what you send it. Pick a provider you trust, or run locally so nothing leaves the room.