Multisis does not lock you into one AI company. Each seat in the council (three Thinkers and one Judge) can use a different provider, so you decide which models do the thinking. This page lists everything you can connect today.
The short version
| Provider | Best for | Key starts with |
|---|---|---|
| OpenRouter | Many free models behind one key | sk-or- |
| Google Gemini | Generous free tier, reads images | AIza |
| GitHub Models | Easy start with just a GitHub account | none |
| Hugging Face | Open models, small monthly credit | hf_ |
| Cerebras | Very fast answers | csk- |
| Groq | Very fast answers | gsk_ |
| Mistral | Free experiment tier | none |
| NVIDIA NIM | Free hosted open models | nvapi- |
| Custom | Any OpenAI-compatible API | varies |
If you paste a key that doesn't match the usual prefix, Multisis will flag it so you can double-check before saving.
The providers
OpenRouter
One key gives you access to dozens of models marked :free. It is the easiest way to try many different models without signing up for several services. Keys are created at openrouter.ai/keys.
Google Gemini
Free access through Google AI Studio with a generous free tier. Gemini models can read images, so this is a good pick for any seat that needs to look at pictures. Get a key at aistudio.google.com/apikey.
GitHub Models
All you need is a GitHub account. Create a token with the Models permission. Daily limits are tight, and Multisis caps output and input length for this provider to stay within them, so it works best for one or two seats rather than all four. It can read images.
Hugging Face
A free token with a small monthly credit. When you create the token, give it the Inference Providers permission or requests will be rejected. Image support depends on the model.
Cerebras
Extremely fast inference with free daily tokens. Cerebras models can't read images, so a seat using Cerebras answers from text only.
Groq
Also very fast, with a free tier that has per-minute limits. If you use Groq for several seats at once, you may hit those limits, so mixing it with other providers helps.
Mistral
Mistral's free "Experiment" tier on La Plateforme. Keys come from console.mistral.ai.
NVIDIA NIM
Free hosted open models from build.nvidia.com.
Custom (OpenAI-compatible)
Any service that speaks the OpenAI chat completions format works, for example SambaNova. Enter the base URL along with your key. For safety, custom URLs must use https and point to a public address. Local addresses such as localhost are blocked on the hosted version.
Mixing providers
You don't have to pick just one. Giving each seat a different provider has real advantages:
- Rate limits spread out. Each provider counts its own requests, so one busy provider doesn't slow the whole council.
- More varied thinking. Different model families approach a problem differently, which gives the Judge more to work with.
- Outages hurt less. If one provider is down, the other seats keep