A study assistant for your own course files. You add your notes, slides, readings and spreadsheets; you ask questions in plain words; Anchor answers from your files only, quoting the exact words and showing where they came from.
PDFs, PowerPoint slides (.pptx), Word documents (.docx), notes (.txt, .md), pictures (.png, .jpg, .webp) and spreadsheets (.xlsx), up to 25 MB each. Drag them into the chat, use the + button, or scan printed pages with the camera button. Pictures inside Word and PowerPoint files aren't read yet. More on uploads.
Yes. Anchor runs in your phone's browser, on Android and iPhone. There's nothing to install.
Yes, printed pages. Tap the camera button in the chat and take a photo of each page; Anchor crops and straightens them, flags any that look blurry or empty, and reads their text by OCR (text recognition). Tap "Check text" under a page to see what was read beside the photo and fix any word it got wrong, then save: the pages become one PDF in your Library. A scanned PDF from your phone's scanner app works too (it is read when added, without the check). Any answer that rests on a scanned page shows that page beside it, and says so if you corrected it. Handwriting isn't read reliably yet. A scan counts as one file. More on scanning.
Lay the page flat on a darker surface, so its edges stand out, and fill the frame with it. Use good light, keep your phone's shadow off the page, and hold still. One photo per page; a scan can have up to 50 pages. If a page is flagged as blurry or empty, tap Retake. If the page can't be found in the photo, it is kept as taken, which still works for a straight, close shot. Before saving, tap "Check text" on each page: bullet points and symbols are the most common misreads.
Nothing is kept until you tap Save, and then only the straightened pages, which are re-saved without the photos' location data. Anchor's own server straightens the photos and reads their text, and the text you check and correct is what is saved. Once saved, the pages are one PDF in your private Library, and deleting it removes them everywhere. As with any file, the passages an answer uses are sent to OpenAI to write the answer, and any graph or diagram on a page is described by OpenAI's model.
Only from your own files. For each question Anchor searches your Library by words and by meaning, picks the best passages, and the AI model answers from those passages alone. Every step of the answer quotes the words it rests on.
That's what it's built for, and we measure it. A made-up answer, or hallucination, is any claim your files don't back, even a true one. On 21 practice questions about a sample course, answered by the same AI model:
These are practice questions, one run each (18 October 2026). The final test on held-back questions hasn't been run yet; we'll publish it when it has.
Each step ends with a Source button. It opens the excerpt beside your chat with the quoted words marked, and “Open in the file” takes you to the page, slide, line or cell. The Sources panel always shows what the latest answer was read from; you can hide it to focus on the chat.
Before an answer is final, every step is checked: its quote must appear word for word in your files, and the step must follow from it. If a check fails, the answer is still shown, marked Unsure, with the check that failed, so you know which part to verify.
Anchor says “Not found in your materials” instead of guessing. If your files answer only part of it, you get that part plus a plain note about what they don't say.
An AI model writes a description of each picture: what it is, all its text, a graph's axes and lines, and its arrows (read three times, keeping what the readings agree on). That description is what answers quote. Because a computer's reading can be wrong, any answer that rests on a picture shows the picture beside it, so you can check.
Yes, less often, and usually visibly. The checks test each step against your files, not arithmetic the model does itself (adding up dates, for example), a picture's description can misread it, and so can the text recognition on a scanned page. Use the Source button on anything that matters.
Only you. Every Library is private; nothing you add is shared with other users, even when classmates add the same file. To read your files and answer your questions, their text and pictures are sent to OpenAI, the AI provider Anchor uses. Questions your files couldn't answer ("Not found" or Unsure) can be shown to the course's Admins, without your name, so the course can fill the gaps. The teaching team can answer them in class and record the answer, which Anchor then gives, cited, to anyone who asks the same thing. Your name is shown with a question only if you switch on "Share my name" under the question box, for class participation.
Yes. Deleting a file removes it from every search index and its text from your saved answers' sources. It can't be used in an answer again.
OpenAI's GPT-5.6 Luna for answering and checking, and OpenAI's text-embedding-3-small for search. The measurements above use the same model with and without Anchor.
Not yet. Connectors are coming with Pro. See the plan.
Standard is S$10 a month, and free for SMU students. Start with a free trial. Pro is S$30 a month and adds connectors and more uploads. Businesses and schools can call 98485438 or email contactus@anchor.com for a proposal. See pricing.