An AI wrapper is a product that sends your input to an AI model someone else built, gets the answer back, and shows it to you inside its own app. That is the entire literal meaning, and by itself it says nothing about whether the product is any good.
You have probably met the phrase in its other use, as a put-down. Somebody posts a new AI tool, somebody else replies "it's just a wrapper," and the thread moves on as though the product has been disproved. That reflex is right about a lot of tools and badly wrong about others, and the difference is testable in about five minutes.
The literal definition, without the sneer
In software, a wrapper is old and boring. It is code whose main job is to call other code, adapting the interface and adding a bit of convenience on the way. Standard programming libraries are full of wrappers around lower level system calls, and the whole point is that the caller gets a cleaner interface without rewriting what already works.
Applied to AI, it means the intelligence does not live in the app. The model runs on the servers of a model provider such as OpenAI, Anthropic or Google. The app you are holding packages your question, sends it over the internet, waits, and renders the reply. Almost every AI product on your phone that is not made by a model lab works this way, including many you would never call a wrapper.
- The app writes the instructions the model actually receives, which you never see.
- It decides what extra context to attach: your files, your history, the current date, a set of rules.
- It handles accounts, payment, rate limits and errors when the model is down.
- It formats the raw text into something readable, and often filters or checks it before you see it.
- It pays the model provider per request, and charges you a subscription that has to cover that.
Why 'just a wrapper' turned into an insult
The insult version of the phrase makes a specific accusation: that the product creates no value of its own and is simply collecting rent between you and a model you could use directly. If the model does all the thinking, the argument goes, the app is a toll booth on a road it did not build.
There is a second fear underneath it, and it is a reasonable one. Products built as a single prompt on top of a general model can be erased overnight when the model provider ships the same feature natively. A tool that only summarises PDFs stops having a reason to exist the week the chatbot you already pay for learns to read PDFs. Investors call this platform risk. Users experience it as an app they paid for quietly becoming pointless.
The third reason is aesthetic, and it is the weakest one. Building on someone else's model looks easy, and things that look easy attract contempt from people who value difficulty for its own sake. Difficulty is not the same as usefulness.
"Wrapper" describes where the computation happens. It does not describe how much work the product does, how much of your life it holds, or whether it saves you anything. Those are separate questions, and only they predict whether the product is worth paying for.
When the criticism is fair
The criticism lands when the app contributes nothing except a text box and a price. These products are genuinely common, especially in app stores, and they share a recognisable profile.
- One prompt, one screen. The product is a template with a marketing name attached to it.
- It remembers nothing. Close the app, reopen it, and you are a stranger again.
- It never touches anything of yours. No files, no calendar, no email, no photos, no documents.
- The paid tier exists only because the free model has a message limit, not because the app does more.
- It cannot tell you what it does when the model is wrong. There is no checking layer, because there is no layer.
- The privacy page is a template, and cannot say where your text goes or how long it is kept.
If a tool matches most of that list, the criticism is not snobbery. You are paying a markup for a prompt you could type yourself, and the moment your usual assistant adds the feature, you will stop opening it.
When the criticism is lazy
"You could build that yourself" has one of the worst track records in the history of software opinions. The canonical example is on Hacker News from 2007, where the top response to the launch of Dropbox opened with a list of qualms, the first of which read: "For a Linux user, you can already build such a system yourself quite trivially by getting an FTP account, mounting it locally with curlftpfs, and then using SVN or CVS on the mounted filesystem." Every word of that was technically correct.
The error is treating the model call as the product. In practice the model call is the cheapest part. The expensive parts are getting your data in reliably, keeping it, retrieving the right piece of it at the right moment, handling the formats nobody wants to handle, staying up, and being simple enough that a person actually uses it on a Tuesday morning. None of that is glamorous, and all of it is the difference between a demo and a tool.
There is also a plain economic point. The model providers sell capability by the token to anyone with a credit card, which means capability is not a moat for anyone, including them. What varies between products is everything built around the call.
The four question test
Skip the label and interrogate the product. Four questions separate a thin reskin from something that earns its subscription, and you can answer all four without reading a line of code.
1. Does it hold durable state about you?
Ask whether the product knows anything today that you told it three weeks ago, and whether that knowledge survives a new chat, a new device and a reinstall. Memory that persists is expensive to build and expensive to run, which is exactly why thin products do not have it. It is also the thing that makes a tool feel like yours instead of like a public utility with your name on the login.
2. Does it do work the base model cannot do alone?
A chat model, on its own, receives text and returns text. It does not go and fetch your bank statement, read the handwriting on a photographed prescription, transcribe a two hour recording, sit inside your calendar, or check its own answer against a source. Every one of those capabilities has to be built and paid for by the product. Ingestion, retrieval, integrations and evaluation are the four places real engineering usually hides.
3. Would it survive the underlying model changing?
This is not hypothetical. Model providers retire models on published schedules, and once a model is retired, requests to it simply fail. Anthropic, for example, publishes a deprecation table with retirement dates and commits to at least 60 days of notice before retiring a publicly released model. Any product that lasts has to be able to swap the engine underneath.
So ask what would be left if the vendor changed. If the answer is your accumulated documents, your history, your connections and your workflows, the product owns something durable. If the answer is nothing, you were paying for access to a model you can access anyway.
4. What actually happens to your data?
Products that pass your text to an outside model inherit that model's data terms, and those terms are public. OpenAI states that data sent to its API is not used to train its models unless you opt in, and that abuse monitoring logs are retained for up to 30 days unless longer retention is required by law or is reasonably necessary to prevent harm to its services or to a third party. That is the floor. What the app itself keeps, for how long, and whether you can delete it is a separate question, and a serious product answers it in specifics.
Run the test in five minutes. Open the app, tell it one specific fact about your situation, force close it, reopen it and start a fresh chat, then ask about that fact. Next, give it something a bare chat model cannot handle: a photographed receipt, a long voice note, a scanned form. Finally, open its privacy page and search for the words retain, delete and train. Products that fail all three checks are the ones the insult was invented for.
Thin reskin versus a real layer
| Signal | Thin reskin | Product with a real layer |
|---|---|---|
| Memory of you | Starts from zero every session | Keeps what you gave it, across devices and sessions |
| Your own material | You paste text in by hand | Ingests files, images, audio, and reads them for you |
| Answer checking | Whatever the model said, verbatim | Cites or returns the source it used |
| If the model changes | Behaviour changes or the app breaks | Engine swaps, your data and workflows stay |
| Data terms | Generic template, no retention detail | Names storage, retention, deletion and isolation |
| What you lose by leaving | Nothing, the model is available elsewhere | Your accumulated library and its history |
The uncomfortable part: nearly everything is a wrapper
Take the strict definition seriously and it swallows the industry. Your banking app wraps a core banking system it did not write. Your ride hailing app wraps mapping and payment services built by other companies. Most business software you have ever used wraps a cloud provider's storage and compute. The browser you are reading this in wraps the operating system's networking and rendering calls.
This is not a gotcha, it is how software has always been built. Abstraction layers exist so that each layer can do one thing well and hand the rest downward. The interesting question was never whether a product calls something else. It is whether the layer it adds is worth its price, and whether it holds anything that would hurt to lose.
Which is why the label predicts almost nothing on its own. There are wrappers that are worthless and wrappers that people would riot over losing. Use the four questions, not the word.
Applying it honestly to MemX
MemX sits in this category, and pretending otherwise would be silly. It does not train its own model. It calls one, like nearly every consumer AI product that is not made by a model lab.
What sits around that call is the actual product: it takes in your photos, PDFs, voice notes and screenshots, runs OCR on the images and transcription on the audio so the text exists in the first place, keeps a persistent per user memory so the things you saved in March are still there in July, retrieves the relevant piece when you ask a plain language question, and returns the source document alongside the answer so you can check it. On data, MemX is private by architecture: per user isolation, customer managed encryption keys, encryption at rest, and capture on your device. Run the four questions on it the same way you would run them on anything else, and if it fails one, that is a fair reason not to use it.
What to take away
"AI wrapper" started as a neutral technical description, became an insult, and along the way stopped carrying information. The honest version of the criticism is narrow and useful: do not pay a subscription for a prompt. The lazy version treats the model call as the whole product and is contradicted by two decades of software history.
The next time you see a new AI tool and the word comes to mind, do not settle the question with the word. Ask whether it remembers you, whether it does something the model cannot do by itself, whether it survives its supplier changing, and what happens to your data. Those four answers tell you everything the label pretends to.
01What does AI wrapper mean?
It means an app that sends your input to an AI model built by another company, then displays the response in its own interface. The app supplies the design, the instructions, the data handling and any extra features. The intelligence itself runs on the model provider's servers.
02Is being an AI wrapper a bad thing?
Not by itself. Most software is built on top of something else. The label only becomes damning when the app adds nothing beyond a text box: no memory, no access to your files, no checking, and no reason to stay if the model behind it changes.
03How can I tell if an app is just a ChatGPT wrapper?
Check three things. Does it remember what you told it last week? Can it handle a photo, a scan or a voice note rather than pasted text? Does its privacy page name retention and deletion specifics? Failing all three is a strong signal of a thin reskin.
04Will AI wrappers get replaced by the model companies?
Single feature ones often are, because a general assistant can add that feature natively. Products that hold your accumulated data, integrations and history are harder to displace, since the value sits in what you built up rather than in the model call itself.
05Is ChatGPT a wrapper?
No, in the strict sense. OpenAI builds the models ChatGPT runs on, so the intelligence and the interface come from the same company. Apps built on the OpenAI API are wrappers around those models, which is a different position in the stack.
