From day one, Correctify has run on your API key: you paste your key, the app recognises the provider from its prefix, and your text goes straight there without passing through our servers. That was simple, and it stays that way. But it was also a narrow door: three providers, one model each, and no say in the matter.

Version 1.9.0 opens that door. The options of the Windows and macOS app now show two drop-down lists: the AI provider and the model. Mistral and OpenRouter join in, and above all you can now run the AI on your own computer, with no key and without a single line of text leaving your machine.

If you don’t want to change anything, don’t

Let’s start with what matters to most people. Google Gemini stays selected by default, with its most affordable model, the very one used until now. If you have already pasted a Gemini key and Correctify suits you as it is, this update asks nothing of you: you will simply see two extra lists in the options, already filled in correctly.

Automatic detection has not gone anywhere either. Paste a key and the provider selects itself: “AIza…” for Google Gemini, “sk-ant-…” for Anthropic, “sk-…” for OpenAI, “sk-or-v1-…” for OpenRouter. What is new is that you can now override it and pick for yourself. And once you do, your choice is no longer overwritten by the detection.

Why this change, and the grain of sand that triggered it

Two reasons, one of them rather unpleasant.

The first is a user request. It came back in many forms, and one of them summed the need up nicely: “I would use several solutions: Ollama and LM Studio locally, OpenRouter and Mistral as providers, but also custom URLs such as LiteLLM or ones from my own projects.” Hard to answer that with three key prefixes hard-coded into the program.

The second is a real incident. Google retired an older Gemini model, overnight, for recently created API keys. The result: users whose every correction failed with a cryptic message, and only one way out, shipping a new version of the app. Fixing a model name should never require recompiling a program and waiting for everyone to update.

That is why the list of providers and models no longer lives inside the app, but on this site. The app downloads it at every startup, and every time you open its options. When a provider retires a model, removing it from the list here is enough: it disappears for everyone, with no new version and nothing to install.

The safety net: a model that vanishes no longer breaks anything

And what if your model is the one that just got retired? The app notices at the next startup and puts you back on your provider’s default model automatically. You see nothing, other than corrections carrying on as usual. That is exactly the Gemini scenario described above, except this time it sorts itself out.

This safety net has one deliberate exception: local servers. The list of models installed on your machine is nobody’s business but yours, and our site knows nothing about it. If you run Ollama with a model we have never heard of, nothing will come along and swap it for another.

The available providers

Google Gemini remains the recommended starting point, because it is the only one with a free tier genuinely generous enough for everyday use. Nothing changes there.

OpenAI and Anthropic (Claude) were already supported and still are, billed as you go, for a fraction of a cent per correction with their smaller models.

Mistral AI joins the list. It is the French provider, for those who would rather use a European model. One practical quirk: its keys have no recognisable prefix, so you have to select it by hand in the “Provider” list. Which is precisely what the list now makes possible.

OpenRouter joins too. It is a middleman: one key, one account, and access to models from dozens of providers, each with its rate on display. It is the cheapest way to try several models without opening an account everywhere. A few models there are even free, with limited quotas.

The API key guide walks through each of them step by step, with ballpark costs.

Running the AI at home, with no key and no bill

This is the most interesting change for anyone who cares about privacy. The app can now talk to an AI server running on your own computer. In that case there is no key, no account and no bill, and above all: your text never leaves your machine. No third-party provider, no terms of service to read, nothing going out to the internet.

Ollama is the most common tool for this. You install it, download a model with a single command, then pick “Ollama (local)” in Correctify. The app queries your server and offers you the list of models you have installed: nothing to type.

LM Studio does the same with a graphical interface, friendlier when you are starting out. Turn on its local server, select “LM Studio (local)”, and you are done. If you changed the port, just fix the suggested address.

Let’s be honest about the trade-offs: you need a reasonably powerful computer, correction is usually slower than with an online service, and quality depends entirely on the model you install. A small local model handles spelling very well, but it will be less comfortable with a subtle rewrite. Up to you to weigh what matters most for your texts.

Your own address: LiteLLM, vLLM, your projects

The last entry in the list is “Custom server”. You enter the address of your OpenAI-compatible gateway, and Correctify fetches the list of available models from it. That covers LiteLLM, vLLM, an Ollama instance on another machine on your network, or the API of a project of your own.

One point deserves to be stated plainly: that address never leaves your computer. It is not sent to the site, and for good reason, it is often an internal address that would make no sense anywhere else. That is also why this entry is the only one you configure entirely yourself, without going through our list.

Picking a model, in practice

Within a single provider, models differ mostly in a trade-off between sharpness and cost. The small model is fast and costs a fraction of a cent; the large one is more subtle and costs several times more. Correctify picks the most affordable one of your provider by default.

That default is not penny-pinching: for fixing spelling, grammar and syntax, small models do the job very well, and the difference with a large one is close to invisible. It becomes noticeable on the rewriting tones, where a more powerful model writes with more nuance. If you rewrite a lot, try a bigger model; otherwise, keep the default.

One key remembered per provider

A small detail that changes everything when comparing: your key is remembered separately for each provider. You can switch from Gemini to OpenAI to try it out, then switch back, without pasting anything again. Keys stay on your device, as before, and are never sent to Correctify.

What about the Chrome, Firefox and Thunderbird extensions?

They do not change. Everything above concerns the desktop app for Windows and macOS. The browser extensions keep working with a Google Gemini, OpenAI or Anthropic key, recognised automatically, with no model choice. This is not an oversight: they are different programs with different constraints, and it seemed better to say so than to let you guess.

By the way, someone joined in during corrections

This version also brings Corry, the little mascot you meet on this site, into the indicator shown while your text is being corrected. He thinks along with you, then waves when it is ready. Purely decorative, and switchable off in the options if you prefer the discreet indicator from before.

How to get it

If Correctify is already installed, there is nothing to do: the app updates itself in the background and lets you know once the new version is in place. Then open the options to find the two lists.

If you are new here, head to the download page, and follow the API key guide to get one for free in a few minutes. Or, if you already have Ollama on your machine, skip that step entirely: you will not need a key at all.