Foundation
Already paying for an API? Run all of MindMax on it
Every AI feature works on your own API key — the same tiers, the same tools, no account and no weekly cap. Point each kind of work at whichever model you want, including one running on your own machine. The subscription is there to save you the setup and the separate bills, not to unlock the app.
- Providers
- 15 pre-filled · any OpenAI-compatible
- Needs
- A key — no account, no weekly cap
- Runs on
- macOS (Apple silicon) · Windows
- Price
- Free app · AI plans from $10/mo, or your own key
- AI
- Included, or your own API key
What bringing your own key usually means, and what it means here
If you have pasted a key into another app before, you probably got a cut-down version of it. Here is the difference — including the row where the simpler approach wins.
| Topic | A typical key boxHow most apps support your own key | In MindMax |
|---|---|---|
| Where the key goes | One box, one model, and everything is sent to it. | Four tiers — quick, everyday, deep, reading images — each pointed at whichever provider and model you choose. |
| Which features work | Often a subset, with the rest kept for subscribers or behind an account. | All of them. No account, no weekly allowance, nothing reserved for subscribers. |
| The model list | Typed by hand, or a dropdown the app updates when it gets round to it. | Fetched from the provider when you open the form, so new models show up the day they ship. |
| Testing it | Checks that the address answers. | Sends one real request through the same path the app uses for real work. |
| A model on your own machine | Sometimes possible, if you find the right address. | Ollama and LM Studio pre-filled, with no key needed. |
| Using the subscription as well | Usually one or the other. | Either one per tier — your key for some kinds of work, the subscription for others. |
| Setting it up | Paste one key and you are done. | More to look at: four tiers to fill. Sensible picks are proposed from the model names, but they are yours to check. |
Each kind of work gets the model you chose for it
Work is not uniform, and neither is pricing. Tidying a paragraph and checking a whole argument should not have to go to the same model.
- The same four tiers the subscription uses — quick work, everyday work, the passes where being right matters, and reading what is on the page rather than in the text — except you decide which model sits in each.
- For example: a cheap, fast model for quick edits; the provider you already use for drafting and reading; the reasoning model you trust for critique passes; and a model that reads images for scans and boxed regions — from a second provider if your main one has none.
- Individual jobs can override the tier, and swapping a provider out later is changing where a tier points, not reconfiguring the app.
A key, not an afternoon of configuration
The usual version of this feature is an empty address box and a dropdown of model names that went stale a year ago.
- The available models are fetched from the provider itself when you open the form, so this year’s models are there on the day they ship rather than when we get round to updating a list.
- Testing a provider sends one real request through the same code path the app uses for everyday work, so a green test means it will actually work — not that an address resolved.
- Your key is encrypted with the operating system’s own credential store before it is written to disk, and it is never sent to us.
What using your own key gets you
The providers you already have
OpenAI, Anthropic, Google Gemini, OpenRouter, Groq, xAI Grok, and DeepSeek, Moonshot Kimi, Alibaba Qwen, Zhipu GLM, Xiaomi MiMo, SiliconFlow — pre-filled with the right address and protocol, so adding one is pasting a key rather than reading API docs. Anything else that speaks the OpenAI API works by pasting its address.
Or a model on your own machine
Ollama and LM Studio need no key at all. Requests go to an endpoint on your own machine, and nothing about them reaches us — see the caveat below about which jobs that is honestly good for.
No weekly allowance
There is no credit meter to watch and no gate that stops a long job halfway. The usage page still counts tokens so you can see what a job actually cost you.
Every feature, not a subset
Reading passes, whole-book translation, book-length drafting, region questions, scanned-page recognition, the canvas and its skills. What a given provider can do is the only limit, and that is covered below.
From a key to a working app
Settings → Providers → add
The list of presets is grouped by where the service is, so you are not scrolling past services you cannot reach.
Pick the one you pay for
Address and protocol are filled in for you, along with a link to the page where that provider issues keys.
Paste the key
It is encrypted with the system credential store before being written to disk. Local providers skip this step — they do not take a key.
The model list arrives on its own
Models are read from the provider as soon as the address and key are in place. If a provider does not publish a list, you type the model name once.
Point each tier at a model
Sensible assignments are proposed from the model names; change any of them, and override per task later when one job deserves a different model.
Where it gets hard
- Your own key is not the same thing as privacy
Pointing a tier at a cloud provider means your document text goes to that provider instead of through us — it does not mean it stays on your machine. The only configuration where nothing leaves the machine is a local model. What is sent in each case is set out on the privacy page rather than summarised loosely here.
- A text-only provider leaves a tier empty
Reading images, scanned pages and boxed regions needs a model that takes pictures, and not every provider has one — DeepSeek says so itself. If your only provider is text-only, recognition and region questions have nothing to run on; adding a second provider for that one tier is the usual fix.
- A local model will not carry the long jobs
Short questions about an open document are fine. Reading a book chapter by chapter, translating three hundred pages or drafting something book-length leans on long context and dependable tool calls, and a model sized to a laptop will be slow at those or will not finish them.
- No allowance also means no brake
The weekly credit limit is what stops a subscription running away with itself. With your own key there is no such limit here — a long job costs what it costs, and the invoice is between you and the provider. The usage page shows tokens spent so it is at least visible.
- You are the one maintaining it
Keys expire, providers retire models, an endpoint changes its address. With the subscription that is our problem; with your own key it surfaces as a job that suddenly fails, and the error says which provider and why.
Common questions
Which providers can I plug my own API key into?
15 providers come pre-filled in MindMax, so adding one is pasting a key rather than looking up an address: OpenAI, Anthropic, Google Gemini and the major Chinese providers among them, plus anything that speaks the OpenAI API. The model list is then fetched from the provider itself rather than read off a list we maintain, so this year's models are there on the day they ship.
Can I run MindMax against a model on my own machine?
Yes. Ollama and LM Studio are among the pre-filled providers and need no key at all — requests go to the endpoint on your own machine, and nothing about them reaches us. Be honest with yourself about the work, though: short questions about an open document are fine, while the long jobs — a book read chapter by chapter, three hundred pages translated, something book-length drafted — lean on long context and reliable tool calls, and a small local model will be slow at them or will not finish them.
Is any MindMax feature locked to the subscription?
No. The subscription is a way to pay for AI compute, not a licence level: with your own key every feature behaves the same way, there is no weekly allowance to run out of, and you do not need an account to use the app at all. What you give up is the convenience — you manage the keys, and the bill from each provider is yours.
Can I use my own key for some work and the subscription for the rest?
Yes. The MindMax subscription shows up in settings as one more provider, and each tier points at either it or one of your own. Credits are spent only on the work that goes through the subscription; everything else is billed by your provider. A tier you leave empty uses your own providers first and falls back to the subscription — if you are signed in — only when none of them fits.
When the subscription is the better fit
Your own key is the right call if you already have one. These are the cases where it is not — and you can switch, or mix the two, at any time.
You would rather not choose a provider per tier
One subscription reaches several model families — GPT, Claude, Gemini, DeepSeek and MiMo — with each tier already assigned. No second provider to sign up for when the first cannot read images.
Plans and weekly allowances →You want a brake
The weekly allowance stops a long job before it runs away, and nothing is charged automatically when it runs out.
You would rather not maintain keys
Expired keys, retired models and moved endpoints are our problem, and there is one bill instead of one per provider.
See also
The step-by-step version, including what to do when a provider publishes no model list.
The other half of not being locked in: your files, in their own folder, unchanged.
If you would rather own the whole pipeline, this is the honest comparison.