Skip to main content
One OpenAI key covers two parts of MagicPost:
  • Writing models draft captions, adapt a post for several networks and write alt text for pictures.
  • A picture model creates new pictures and edits pictures from your media library.
OpenAI has no video model in MagicPost. Use Google Gemini, xAI Grok or fal.ai for clips.
There is one OpenAI key. The OpenAI GPT field on Writing Models and the OpenAI field on Media Models read and write the same stored key.

Before you start

  • You can sign in to the MagicPost admin panel.
  • You have an OpenAI account on the API platform with billing set up. OpenAI bills you directly for every request. This is separate from the credits your users spend in MagicPost.

Get an API key

1

Sign in to the API platform

Go to platform.openai.com and sign in. This is the API platform, not the ChatGPT app.
2

Set up billing

Add a payment method or a credit balance in your billing settings so requests can go through.
3

Open the API keys page

Go to platform.openai.com/api-keys. If your account uses projects, select the project that should own the key.
4

Create the key

Click Create new secret key, name it (for example, MagicPost) and create it. If you restrict the key’s permissions, allow chat completions and image generation and editing.
OpenAI shows the full key only once. Copy it now. If you lose it, create a new key.

Add the key in MagicPost

1

Open Writing Models

Go to Admin → Publishing → Writing Models.
2

Paste the key

In the Provider keys box, paste the key into the OpenAI GPT field and click Save keys. The toast reads Keys saved. and the OpenAI GPT card shows Key stored.
3

Switch on the models

Turn on the switch next to each model you want to offer. Click Test on one of them. MagicPost sends a one-word request and shows either that the model answered or the error OpenAI returned.
4

Check the picture model

Open Admin → Publishing → Media Models. The OpenAI field already holds the same key and the card shows Key stored. Switch on the picture model and click Test. MagicPost asks for one small picture, so the test makes a real, billable request.
Test works whether or not the model’s switch is on, so you can check a model before users can see it.

Starting models

The installer loads this list. Identifiers are as shipped and can be corrected later (see Model identifiers). Credits are MagicPost credits, not OpenAI prices, and you can change them on the model. Writing Models Media Models

How MagicPost calls OpenAI

  • Writing uses OpenAI’s chat completions endpoint. Models that reason (identifiers starting with o and a digit, or gpt-5 and later, except -chat variants) are sent a low reasoning effort and extra room in the token limit, and no temperature.
  • If OpenAI rejects temperature or reasoning_effort by name, MagicPost sends the request again without it and remembers that for the model. The Test result then reads that the model turned down those settings, so they are no longer sent.
  • Pictures use OpenAI’s image generation endpoint, or its image edit endpoint when the user starts from a picture. Wider-than-tall requests are made at 1536x1024, taller-than-wide at 1024x1536 and square at 1024x1024.
  • Attaching pictures to a caption request needs a model that can look at pictures. MagicPost recognizes this from the identifier. If a model cannot, users see that the model cannot look at pictures and are asked to pick another.

Troubleshooting