Skip to content
CouverturesCouvertures

Setting up an AI

Create a key at Anthropic or OpenAI, or install Ollama on the Mac, step by step.

An AI is not required: the barcode is enough for most books. It helps with books without a barcode, whose cover it reads (or a movie’s case, an album’s sleeve), and to understand dictated titles. It also reads the contents of magazines, and suggests keywords (in To review and in the record, with a tap on the sparkles, and in Classification for books already there). You set it up in Settings, “Cover recognition”: on the Mac, or on the iPhone when the library lives there.

Without a key, Apple Intelligence already reads covers, on the device, if it has the model (below). For a surer AI, three choices:

  • Claude (Anthropic): the most reliable. A few cents per hundred books, paid in advance.
  • OpenAI: the same, at OpenAI.
  • Ollama: free, and the photo never leaves your Mac. Slower, and it needs a recent Mac; it is for a library that lives on the Mac.

For Claude and OpenAI, the key comes from an API account, with a little credit. A Claude or ChatGPT subscription is not enough: it is a separate account, paid separately.

Without a key: Apple Intelligence

As long as no key is saved (and “No AI” is not chosen), Couvertures hands the photo of a cover to Apple’s model, if it is there: an iPhone, an iPad or a Mac on iOS 27 or macOS 27, with Apple Intelligence turned on. The reading happens on the device: the photo does not leave it. The model only reads covers, and a disc’s sleeve: not magazine contents, nor keywords, which need a key.

Its reading is checked against the text the device reads itself on the photo: misread words are corrected, names that are not there are dropped. Review says “Your photo, title read by Apple Intelligence”; the proposal is always to be checked, and the words the device did not find on the photo are listed under the title: “Check on the cover: …”. The title can be retyped as usual. If the model is busy, the card says “Analysis failed”: tap Retry the analysis a little later.

Without that model, Review simply says “Type the title”: type it, Couvertures looks for the rest. As soon as a key is saved, these photos start again by themselves.

Claude

  1. Open platform.claude.com and create an account.
  2. In Billing, add a card and buy credit. The minimum, 5 dollars, is enough for thousands of covers. Without credit, the key is refused.
  3. In API Keys, choose Create Key, name it “Couvertures”, then copy the key (it starts with sk-ant-). It is shown only once: otherwise, create another one.
  4. In the Settings of the Mac (or of the iPhone, if the library lives there), choose Claude, paste the key into “Anthropic API key”, then Test the connection.

OpenAI

  1. Open platform.openai.com and sign in. Your ChatGPT account works, but its subscription does not pay for the API.
  2. In Settings, then Billing, add a card and some credit: 5 dollars is enough.
  3. In API keys, choose Create new secret key, name it “Couvertures”, then copy the key (it starts with sk-). It is shown only once.
  4. In the Settings of the Mac (or of the iPhone, if the library lives there), choose OpenAI, paste the key into “OpenAI API key”, then Test the connection.

OpenAI does not use the photos to train its models, and keeps them for up to 30 days.

Ollama

You need a Mac with Apple silicon, preferably with 16 GB of memory or more.

  1. Download Ollama from ollama.com, drag it into Applications, then open it. Its icon appears in the menu bar.
  2. Open Terminal (in Applications, then Utilities), type the following line, then Return: ollama pull qwen2.5vl:7b. The model weighs about 6 GB; wait until it finishes.
  3. On the Mac, in Settings, choose Ollama. The address and the model are already filled in: Test the connection.

Ollama must be open when Couvertures reads a cover.

If the test fails

  • Key refused: copy it again in full, with no space before or after. Or create another one.
  • No credit left: add some, in Billing.
  • Ollama unreachable: open the Ollama app.
  • The model cannot read images: you need a vision model, such as qwen2.5vl:7b; a text model, such as llama3.1, cannot read covers.

These sites sometimes rename their menus. Their own help remains the reference: Anthropic documentation, OpenAI help, Ollama documentation.