Why the AI key should be yours

Almost every AI content tool sells the same promise: describe your brand, receive finished posts. The promise is a good one. The problem is a decision the tool makes quietly, the moment you subscribe — the decision of which model will generate your content, in which version, at what cost.

That decision looks technical and unimportant. It is not. It defines what you pay, what you get, and how much of it you can fix when it comes out wrong.

1. Model lock-in is a lock-in you cannot see

When the AI vendor is embedded in the tool, you do not know which one it is. You do not know the version. You do not know when it changes. And it does change: vendors deprecate models, raise prices, swap the default from one version to the next.

The classic symptom is this: for three months the captions came out in your voice. On some random Tuesday, they start coming out longer, more generic, with that rhythm of AI text everyone now recognises. You changed nothing. Your brand changed nothing. The model underneath changed — and you have nowhere to look, because the tool does not show what it used, nor does it have a place where you could say “go back to the previous one”.

Compare that with the case where the provider is a record you own. The model is a field. If the output gets worse, you change the field. If you want to test two models on the same piece, you create two providers. If a vendor goes down, you point at another one without waiting for a status page announcement.

2. You pay for the generation twice

A tool with embedded AI has two ways of charging, and it normally uses both: a monthly subscription and a generation cap inside it. The cap is sold as credits, or as “posts per month”, or as a number of images. What it actually is, in practice, is the vendor cost plus a margin — on a cost you cannot see.

There is nothing scandalous about charging a margin. The problem is the asymmetry: you cannot judge whether the margin is fair, because you do not know the cost. And unlike almost every other software input, this cost is public. AI vendor prices are listed on their own websites. Anyone can work out what a thousand captions cost — anyone except you, because you do not know which model generated yours.

When the key is yours, that arithmetic closes in front of you. The vendor invoice arrives itemised by call. You see that the text cost almost nothing and the video cost everything else — and you start deciding with information: maybe it is worth generating video for fewer pieces, or using a cheaper model for drafts and the expensive one only on the piece that goes live.

3. The credit you already bought does not work

This is the most irritating of the three, and the most common.

Plenty of people who subscribe to a content tool already have an account with OpenAI, Google or another vendor. They already pay per use. Sometimes they have promotional credit about to expire. None of it works inside the tool, because the tool only knows how to use its own account. You end up paying two AI bills in the same month for two similar things.

For anyone with a corporate account under a data agreement it is worse: the company has already negotiated terms with a vendor, already validated where the data lives, already been through legal review. And the content tool sends everything to a vendor nobody reviewed, over a path nobody approved.

What “bring your own key” solves — and what it does not

BYOK is the acronym for this idea. It solves the three problems above directly: the model is your choice, the cost is the vendor cost with no middleman, and the account you already have starts counting.

It does not solve everything, and it is worth saying what:

• You have to create the account at the vendor. Ten minutes, once. But they are ten minutes the alternative does not ask for.
• The error becomes yours to resolve. If the key expires or the quota runs out, you are the one who has to see the message. That is why the per-provider health test returns the raw vendor message instead of a “generation error” — without it, BYOK becomes guesswork.
• Not everyone wants this. Someone running a bakery should not have to understand what an API key is in order to publish a story. Which is why the alternative has to genuinely exist, and not as a punishment plan.

That is why both paths exist in the same EverFeed plan. You plug in your key and pay the vendor directly, or configure nothing and use our credits. What does not exist is the third option — the one where you cannot choose.

How this looks in the product

The idea is only worth something if it reaches the product, and in EverFeed it arrives as a record. A provider is: name, capability (text, image or video), driver, base URL, key, default model, list of models, a free-form parameters field in JSON, timeout, whether it is the default, enabled or disabled.

Three consequences fall out of that almost for free:

One vendor per capability. Because each pipeline step asks for a capability, nothing forces text and video to come from the same place. Text from a cheap fast provider, images from another, video from a third. The three decisions are independent.

A vendor that is not on the list gets in anyway. There is a driver that describes an entire API in JSON — authentication, text route, image route, video route, health route. Anyone with a model running on their own infrastructure, or a corporate gateway, plugs in without waiting for a release from us.

The key does not come back. It is stored encrypted, and the API never returns it: the panel receives something like sk-…4f2a and nothing else. That is not marketing comfort — it is the shape of the JSON our own interface receives. Not even it can read the whole value back.

The question that remains

If you are evaluating any AI content tool, including this one, ask three questions before subscribing:

1. Which model generated this piece? If the answer is not a name and a version, the tool is hiding the most expensive decision it makes on your behalf.
2. Can I use my own account? If not, you will pay twice for the same capability.
3. Where do I see what was sent? Without the exact prompt, “improving the result” becomes paid trial and error.

They are questions any honest tool answers in a minute. The ones that do not answer have already answered.