Home / Library / Help / Set up a runner
Guide
Set up your model runner
mi·do·na doesn't ship a model of its own. You run one and point mi·do·na at it. A runner is the app that loads a language model and serves it over a local API on your machine, so nothing you send ever leaves your computer. Pick one below, then follow its guide.
Pick a runner
All four run models locally and expose an OpenAI-compatible API that mi·do·na can reach. They differ mostly in how much they do for you versus how much they hand you the controls.
- LM StudioA polished desktop app with a built-in model browser. The easiest place to start if you'd rather click than type.Local
- OllamaCommand-line first and minimal, with sensible defaults. One command pulls a model, and the server runs itself in the background.Local
- llama.cppThe inference engine most other runners are built on. Maximum control: you manage the GGUF file and the server flags yourself.Local
- koboldcppA single-file executable with a simple launcher. A long-time favorite in the roleplay community.Local
What every runner needs
The details differ, but the shape is the same across all four. Each guide walks these steps for that specific runner.
Download a model
Grab a model file that fits your RAM or VRAM. When in doubt, start in the 7-8B range, then go bigger once it's working.
Load it and start the local server
The runner loads the model into memory and serves it over an OpenAI-compatible API.
Allow browser access (CORS)
mi·do·na runs in your browser, so the local server has to permit browser requests. Some runners allow this by default; others need a setting or an environment variable. Each guide says which.
Connect mi·do·na
Open the connection panel, point mi·do·na at the runner's endpoint, and load models. See Connect a model.
Next: connect mi·do·na
Once a runner is serving a model, head to Connect a model to wire it up.
Served over HTTPS and can't reach localhost? The browser
can block an HTTPS page from talking to http://localhost - most
often in Safari. See
Reaching a local model over HTTPS
.
Not local? Optional alternatives
Don't want to run a model yourself? mi·do·na also connects to cloud providers - OpenRouter (a paid gateway to many models) and AI Horde (a free, community-run GPU pool). These are the trade-off: no setup, but the model runs on someone else's hardware, so they're not local and not private - whoever hosts the model can see what you send. There's also a Generic OpenAI option for pointing at any OpenAI-compatible URL. See Connect a model for all of them.
No wrong choice here. If you're unsure, LM Studio is the gentlest start and Ollama the fastest once you're comfortable in a terminal.