You can attach a file to AI on Mac without opening a chat site. In Summon, add the file from the panel, and the extract goes to the model behind the key you already added.
How do you attach a file to AI on Mac?
Press ⌘K, then add the file. The paperclip opens a file dialog. You can also drop a file on the panel, or paste one. Summon reads it on your Mac before the command runs.
The formats are the ones you already have lying around: plain text, PDF, Word (.docx), EPUB, and images (png, jpg, gif, webp). A PDF is the text PDFKit can read. A Word file is the document text. An EPUB is the chapter text, with the title and author when the file has them. A text file is capped at 400,000 characters. Anything over 50 MB is refused, except an EPUB.
Then run a command. A built-in preset, a command you saved, an instruction you type, and a follow-up in the result panel all use the attachment. The chip stays on the result so you can see what went with the question.
Does the file leave my Mac?
The file itself stays on your Mac. Summon sends the extract, not the original file, to the provider behind your API key: OpenAI, Anthropic, Google Gemini, or xAI. The key stays in the macOS Keychain. Summon’s servers do not receive the file, the extract, or the reply. That is the same path as a command on selected text, described in the privacy policy.
Optional usage analytics never include attachment contents, and you can turn them off in Settings.
Can I attach a screenshot or a web page?
Yes. The viewfinder next to the paperclip opens the system screenshot controls: the whole screen, one window, or a selection. The picture is resized so the longest side is 2,048 pixels and sent as an image, so the model can see it. A model that cannot take images skips the picture and the chip says so. A processed image over 20 MB in base64 is refused.
A copied web address is the other path. Paste a single http or https link into the panel, or select one in the address bar and press ⌘K. Summon fetches that public page, strips the HTML, and sends the text, capped at 300,000 characters. The chip shows the page’s first real image, or its favicon when there isn’t one. A page that is mostly JavaScript comes through thin, and the chip says so.
Which commands work on an attachment?
Any command you run while the chip is on the panel. The built-in presets, a command you saved, and the ones in the command library all receive the attachment. Brief me is the short version of a document. Read this image transcribes a screenshot. Open questions lists what the source never answers. TypeScript interface turns JSON or a schema into types.
You can also type the instruction yourself. An empty field with a file attached sends “Use the attached content.” A follow-up in the same result can ask a second question, and an earlier image is sent again because the provider does not keep it between requests.
What does this not do?
Summon does not OCR a scanned PDF or a photo of a page. If PDFKit finds no text, you get a notice, not a transcription. It does not read a page you are logged into, and it does not pull a YouTube transcript. It does not index a folder of documents or keep a chat you can reopen later. Each command is one run: the extract goes to your provider, the result stays in the panel until you insert it or dismiss it.
Summon runs on Apple silicon Macs with macOS 15 or later.