AI features
Adding AI to an app you built, what it costs, and why building the app and running it are billed separately.
If your app needs a model at runtime, a coach that replies, a summariser, a classifier, image generation, transcription, it runs through Mistflow's AI gateway. You do not bring an API key or choose a provider.
What is available
| Kind | Use for |
|---|---|
| Chat | Conversational replies, assistants, coaching |
| Reasoning | Harder analysis where quality matters more than speed |
| Classify | Tagging, routing, moderation, structured extraction |
| Image | Generating pictures |
| Speech to text | Transcribing voice notes and recordings |
| Text to speech | Reading content aloud |
| Embeddings | Search and recommendations by meaning rather than keyword |
Ask by outcome:
After each run, write a short encouraging summary of how it went
Tag incoming messages as question, complaint, or praise
Let people search their notes by what they mean, not exact wordsTwo balances, never mixed
This trips people up, so it is worth being explicit:
| Balance | Pays for | Who spends it |
|---|---|---|
| Build credit | Generating and changing your app | You, while building |
| AI credit | Your finished app calling a model | Your app's users, billed to you |
Building an AI coach costs build credit. People talking to the coach costs AI credit. The two are kept separate on purpose so that a busy app cannot consume your ability to keep building it, and a heavy build session cannot break your live app.
Costs and control
Runtime AI is metered per call and billed to you as the app's owner. There is a monthly cap so a runaway loop or a burst of traffic cannot produce an unbounded bill.
Recent activity, spend, and the model in use are visible in the dashboard's AI section.
Bringing your own key
If you already have an OpenRouter account, you can point a role at your own key and those calls bill to you there instead of drawing Mistflow AI credit. Keys are stored encrypted and are never used for anything other than your own calls.
Where the key lives
Worth knowing, because it explains the shape of an AI app.
A phone app has no server of its own, and anything shipped inside a phone app can be read by anyone who installs it. So the key that lets your app call a model never lives in the phone app. It lives on your app's paired backend, and the phone app talks to that backend using the signed-in person's own credentials.
The practical consequence: an AI feature needs the app's backend half, which is built for you. It is not something you configure, but it does mean "add AI" is a real feature, not a switch.
Choosing models
You do not pick a model by name. Your app asks for a kind of work, chat or classify or image, and Mistflow routes it to a current model. When a better model ships, apps move across without being rebuilt.
If a specific model genuinely matters, it can be pinned. Most apps should not.
Failures
- A transient provider failure is retried once on a backup model
- A call refused for budget, configuration, or provider reasons comes back with a specific reason rather than a generic error
- Your app should still have an answer for "the model did not respond". Ask for it: "if the summary fails, show the run without one"