Use your own key

Connect Claude, OpenAI, Gemini or Mistral with an API key, or a server such as Ollama, LM Studio, Scaleway or OVHcloud, and choose its model.

Mac · iPhone · iPad

On this page

Instead of Apple Intelligence, or when your device doesn't have it, Courant can use a service you connect with your own key. You pay the service directly, for what you use: Courant has no key of its own and sells no AI credit. Requests go straight from your device to the service.

ServiceWhere to get a keyDefault model
Claude (Anthropic)platform.claude.comclaude-haiku-5-5
OpenAIplatform.openai.comgpt-6.1-terra
Gemini (Google)Google AI Studiogemini-3.8-flash
Mistral AIconsole.mistral.aimistral-small-latest
OpenAI-CompatibleYour server or host (none for Ollama or LM Studio)The one you type

Connect a service

  1. Open Settings › Intelligence and choose the service's card. Its page opens, since it needs a key.
  2. Create a key with Get a … API Key, under the field, and paste it in API Key. For Claude, the page first lists three steps, each with its link: Create an Anthropic account and add credit, Set a monthly spending limit, Create a key and copy it.
  3. As the key is saved, Courant asks Send your emails to …? Choose Allow: until you do, nothing is sent to that service from this device.
  4. Optionally, choose a Model (the chevrons beside the field list the suggested ones), then click Test. A test sends a single word, no email: « Works — answered in 0.6 s », or the service's error.

The page's first lines say where the service runs, what it receives and who bills you, with links to its Privacy Policy and Terms.

Choose the model

Each page suggests a fast, inexpensive model by default and a better writer:

  • Claude: Haiku is fast and inexpensive; Sonnet writes better. With Haiku, a thousand summaries or replies cost about $0.50; with Sonnet, about $8.
  • OpenAI: Terra is fast and writes naturally; Luna costs less; Sol writes better.
  • Gemini: Flash is fast and inexpensive; Flash-Lite costs less; Gemini 4 Argon writes better.
  • Mistral AI: Mistral Small is fast and inexpensive; Mistral Medium writes better.

You can also type the name of any model the service offers. Leave the field empty for the default.

Your own server, or a European host

The OpenAI-Compatible card takes any server that speaks OpenAI's API:

  1. Under Known Servers, tap Ollama, LM Studio, Scaleway or OVHcloud AI Endpoints to fill in its address, or type it in Server (it usually ends in /v1).
  2. Add a key in API Key (Optional) if the server wants one (none for Ollama or LM Studio).
  3. Type the Model, for example llama3.2 with Ollama, then Test.

A server on this device or your local network is yours: Courant doesn't ask permission for it, and your emails stay there unless the server sends them on. A server on the internet asks, like the services above.

On all your devices

The keys, models and server addresses are kept in your iCloud Keychain: enter a key on the Mac and your iPhone has it too. Two things stay per device: the model chosen (see Choose a model), and the permission to send your emails, which each device asks for once. Until then, the card says Needs your permission: open its page and choose Allow Sending Emails to ….

Withdraw, remove or forget

  • Withdraw Permission, on the service's page: Courant sends nothing more to it from this device. The key stays.
  • Remove Key (Forget This Server for a server): it's removed from all your devices, and Courant stops sending emails to that service.

What is sent

When you use a feature, the text of the emails concerned, with their senders and recipients. The automatic summaries stay with Apple Intelligence on the device unless you turn on Automatic summaries with … in Settings › Intelligence. Details in AI and your data.