Scraping
Paste a link or a YouTube video into any prompt and Triplo reads it for you: what gets fetched, how much reaches the model, where links can come from, and what leaves your machine.
What scraping does Paid plan
What it is. Put a link in a prompt and Triplo fetches it before the model sees anything. A web page is replaced by its text, a YouTube video by its transcript. You ask about the content; the model answers with the content in front of it, not just the address.
How to reach it. The bug icon in the bottom bar (tooltip: Web scraping mode), or Cmd+Shift+S while the window is focused. On by default. There is nothing else to set up: no command, no prefix, no special syntax.
Scraping works on the prompt you send, right now, and nothing is kept afterwards. To build a lasting library from web pages that every future prompt can draw on, add them to a Mind instead. That is a different feature, with its own import.
What happens when you press Enter
Every http:// or https:// address in the prompt, once each, in the order you wrote them. YouTube links are recognised and handled separately from ordinary pages.
The turn shows up straight away, with a line saying what is being fetched: Reading example.com… for a page, Fetching the video transcript… for a video. Links are fetched one after another, and the line follows along.
The address is replaced, in place, by what was fetched, so the model can always tell your words from the source's.
With the Instruction, the Mind passages and the conversation, exactly as any other prompt. If a Mind is selected, the status line switches to Looking through your knowledge base… before the answer starts.
This is what the model receives in place of a link:
Stop works during the fetch. Press the stop button or Esc while the status line is showing and the turn is cancelled before anything reaches the model.
Web pages
Triplo downloads the address and keeps the text. What that text looks like depends on what the server sends back:
| The link returns | What the model gets |
|---|---|
| An HTML page | The visible text. Scripts and styles are removed, tags dropped, whitespace collapsed. Menus and footers are text too, so they come along. |
| A PDF | The PDF's text layer, read properly rather than stripped. |
| Plain text, Markdown, JSON, CSV | The content as it is. Often the cleanest source there is. |
Each page contributes up to 3,000 characters, roughly a page and a half of prose. On most articles the part that matters is near the top, so this loses little; on a long reference page it is the first screens only.
Getting cleaner text
- Prefer the raw version. A file on GitHub? Link the
raw.githubusercontent.comaddress, not the page around it. An API with a JSON endpoint? Link that. - Link the exact page. A deep link to the section you care about beats the home page, whose 3,000 characters are mostly navigation.
- Print and reader views help. Where a site offers a print or text-only version, its first 3,000 characters are almost all content.
- Need the whole document? Download it and attach it, or add it to a Mind, which retrieves the relevant parts of a document of any length.
YouTube videos
A YouTube watch page is almost all player and recommendations, so Triplo never scrapes it. It fetches the video's transcript instead: what is actually said, as plain text. This is what makes questions like "what does she recommend at the end?" answerable.
Links it recognises
| Shape | Example |
|---|---|
| Watch page | youtube.com/watch?v=…, also on m. and music.youtube.com |
| Short link | youtu.be/… |
| Shorts | youtube.com/shorts/… |
| Live streams | youtube.com/live/… |
| Embeds | youtube.com/embed/…, youtube-nocookie.com/embed/… |
Extra parameters do not matter: a timestamp (&t=90s), a playlist (&list=…) or any order of them still resolves to the video. A playlist or channel link with no video in it is not a video, and is read as an ordinary page.
How much of the video
Transcripts are not held to the 3,000 characters a page gets, because a few minutes of speech is rarely enough to reason about a video. Triplo sizes each one to the model you are using:
- Up to about three hours of speech on a model with a large context window.
- Less on a smaller model. Triplo measures what the conversation and the rest of the prompt already take, and keeps the transcript inside what is left, so the request is never rejected for being too long. If it does not recognise the model, it assumes a small window and keeps a shorter excerpt.
- Several videos share the space. Two videos in one prompt each get half of what one would.
- A cut transcript is marked. It ends on a whole word with
[…truncated], so the model knows it has not seen the end.
Language
With an output language chosen in the globe menu, Triplo asks the transcript service for the captions in that language. Set to Default, the service picks the track. Either way, the answer follows your output language.
A transcript is long, and in chat mode it travels with every follow-up. Ask everything you want about one video in the same conversation rather than pasting the link again, and clear the chat (Shift+Backspace) when you move on.
Links do not have to be typed
Triplo looks for links in the whole assembled prompt, not only in what you typed. A link counts wherever it comes from:
//selected////clipboard//Recipes
Summarise any link, from anywhere. Create a SmartPrompt such as "Summarise this in five bullets: //selected//" and give it a hotkey. Select a link in an email or a chat, press the key, and the summary arrives without opening the page.
A prompt that always reads the latest version. Put a fixed address in a SmartPrompt: a status page, a changelog, a price list, a JSON feed.
Watch a video in your own language. Set the globe to Portuguese, paste an English talk, and ask for the key points. The transcript is requested in Portuguese and the answer comes back in Portuguese.
When links are sent as they are
In these cases nothing is fetched, and the model sees the address exactly as you wrote it:
| Situation | Why |
|---|---|
| The bug icon is off | Scraping is a switch. Turn it off when you want to talk about a URL rather than its content. |
| Agent mode (Pi) | The agent gets your prompt untouched and decides for itself what to do with the link. |
Raw pass-through (// at the start) | Pass-through skips everything Triplo would add, scraping included. |
| The fetch failed | No connection, a timeout, or a video without captions. The rest of the prompt still sends; see Limits. |
What leaves your machine
| Step | Goes to | What is sent |
|---|---|---|
| Reading a web page | The website | An ordinary request from your machine, made by the app. Your browser's cookies and sign-ins are not used. |
| Fetching a transcript | Triplo's transcript service | The video's ID and, if set, your output language. Never your prompt. |
| Answering | The model you selected | The prompt with the fetched text in it, like any other prompt. |
With a local model, the fetched text stays on your machine once it arrives. For a fully local-only setup, turn scraping off so a pasted link is never fetched.
Limits & gotchas
- Pages behind a login, or built entirely by JavaScript, usually come back empty or as a sign-in screen. Copy the text and paste it instead.
- A failed fetch is quiet. The link stays in the prompt as plain text and the model only sees the address. If an answer reads like the model never opened the page, it probably did not.
- An error page is still a page. A link that returns "not found" or "access denied" sends that text to the model.
- Videos without captions have no transcript to fetch, and the link is left as it is. Triplo never falls back to scraping the watch page. A transcript that takes longer than 30 seconds is given up on.
- Punctuation right after a link can become part of it: "read https://example.com/page." fetches the address with the full stop. Leave a space, or wrap the link in parentheses. YouTube links are tidied automatically.
- One fetch per address. Repeating a link in the same prompt fetches it once, and the content goes where it first appears.
- Every link costs time and tokens. Five links are five fetches, one after another, and their text counts towards your allowance or your provider's bill.
- The chat bubble shows what you typed, not the fetched text. The model gets the full version.