Reads page by page under a budget
Tools force page reads under a character budget. A large lease or report does not flood the first turn with a fake full-coverage summary.
PageWise · Local document agent · macOS / Windows / Linux
See what the agent read, page by page
Local PDF workbench with page-wise reads under budgets. Marks stay on the document; citations bring you back to the source across macOS, Windows, and Linux.
Local PDF workbench
Ask about a page — see what was read
Open a PDF, ask a question, and watch the agent read specific pages under a budget. Marks stay on the document; citations and follow-along preview keep the source beside the answer.
Open or continue
Start with a document on disk
Open a PDF or drag it in. Recent files stay local so you can continue the same thread.
What you get in the workbench
What you get in the workbench
Text layer or vision — same tools
Use the PDF text layer when it exists; use vision when the page is image-heavy. The agent calls the same page tools either way.
The read trail becomes controls
Clickable page citations, pages-read chips, follow-along preview, and ask-about-this on a selection let you jump the preview to what the agent just used.
Workbench around the page
Preview with thumbnails, zoom, in-doc search, and outline or bookmark tree. One chat thread per document, saved locally, with Markdown export.
Marks stay on the document
Highlight and note marks persist per document, show in the preview, stay visible to the agent, and export with Markdown.
How a session runs
How a session runs
Open a document
Pick or drop a PDF or image. PageWise loads the preview and indexes text locally — path permission stays on your machine.
Ask, or select then ask
Type a question, or highlight text and ask about the selection. One thread stays with that document.
Watch the tool loop
The agent plans outline, search, and page reads inside step and read budgets, then writes a grounded answer on the last step — a search miss is not “does not exist.”
Follow and verify
Jump via page citations or follow-along preview so the preview tracks what was read. Decide what to trust with the source still in view.
Local by design
Local by design
Documents stay on disk
PageWise does not upload your PDFs or images to its own servers. Processing stays on your machine.
Only what the model needs leaves
Extracted text — and optional vision payloads you enable — go to the LLM you configure. Nothing else leaves without that setup.
Keys in the OS keychain
Credentials use macOS Keychain, Windows Credential Manager, or Linux Secret Service. Assistant and Scan (vision) can use different models, including Ollama.
Questions & how-to
Can it run fully local with Ollama?
Yes. Point Assistant (and Scan if needed) at an OpenAI-compatible Ollama endpoint. Files stay on disk; only the payloads you enable go to that local model.
When should I enable OpenRouter web search?
Only when a question needs outside verification. It is opt-in per call and off by default, so ordinary document Q&A is not polluted by web results.
Do my files leave my computer?
No upload to a PageWise server. Only extracted text — and optional vision payloads — go to the model endpoint you configure.
How do scanned pages work?
When a text layer is thin or missing, indexing can use a vision / multimodal Scan model. Prefer a vision-capable model for image-heavy PDFs.
What platforms does it run on?
GitHub Releases ship macOS DMG, Windows setup/MSI, and Linux AppImage/deb. Builds are unsigned, so the first open may need an OS exception. The project is MIT-licensed.