The PDF reader that reads with you
A fast, native reader built in Rust. Next, the AI agents you already use will read alongside you: summarizing, finding and citing pages, and highlighting what matters, right on the page.
12. Term and Termination
12.1 Either party may terminate this Lease on the third anniversary of the Commencement Date by giving the other party not less than six months' written notice.
Either side can leave at year three with six months' written notice. p. 4, 12.1
Agent panel preview, in development
Office Lease Agreement
12. Term and Termination
12.1 Either party may terminate this Lease on the third anniversary of the Commencement Date by giving the other party not less than six months' written notice.
12.2 The Landlord may terminate by notice if the Rent is unpaid for twenty-one days after becoming due, or if the Tenant commits a material breach that is not remedied within thirty days.
13. Rent Review
The Rent is reviewed on each anniversary in line with the consumer price index, subject to a cap of four per cent per year.
Fast, calm, and native on your desktop.
No browser engine, no account, no upload. Agentic PDF is written in Rust on the GPU-accelerated GPUI framework, with Google's PDFium engine doing the rendering.
Tabs that keep your place
Open many PDFs side by side. Each tab remembers its page, zoom and scroll position.
AvailableSelect and copy
Drag across real text, including rotated pages. Copies keep Unicode and line breaks intact.
AvailableKeyboard first
Command palette, go to page, zoom and tab switching without leaving the keyboard.
AvailableSearch the whole document
Matches across line breaks and accents, with every result listed by page and highlighted in place.
Next releaseOpen from Finder
Open With and double-click open straight into a tab, even when the app is already running.
Next releasePick up where you left off
Recent files reopen on the page and zoom you left them at.
Next releaseBring the agent you already use. It reads, cites and asks first.
Agentic PDF connects to coding and research agents installed on your computer through the open Agent Client Protocol. No API keys to paste and no new subscription. The reader gives the agent precise tools for your document, and keeps you in control of every change.
Connect your agent
The reader finds the agents installed on your computer and starts one in a panel beside your document.
It reads with real tools
Outline, search, and text with page positions, served by the reader over MCP in small, bounded pieces. Long documents stay fast and memory stays flat.
It answers with pages, and asks before acting
Every claim links to the page it came from. Highlights and notes are only added after you approve them.
Private by default
Your PDF stays on your computer. Nothing leaves unless the agent you chose sends it.
No lock-in
Switch between agents as they improve. The reader never ties you to one model or vendor.
Answers you can check
Page citations jump to the exact passage, so you verify instead of trusting.
You approve every edit
Reading is automatic. Writing to your document always waits for your yes.
Summarize this contract and list every deadline.
A short summary with each date cited to its clause and page.
Where does it say we can cancel?
The exact passages, highlighted on the page after you approve.
Compare section 4 with the appendix.
Differences explained side by side, each linked to its page.
Which pages could you not read?
Scanned pages are named, never skipped silently.
Fast enough that you stop noticing it.
We publish what we measure, with the machine and method. Search runs in the background in short slices, so turning pages never waits for it.
to search all 457 pages of a technical book, from 425 to 620 ms across three runs.
until the first matches appear while the rest of the book is still being searched.
at most to render a new page while a full search is running.
- 0.5 s to search all 457 pages of a technical book, from 425 to 620 ms across three runs.
- 40 ms until the first matches appear while the rest of the book is still being searched.
- 18 ms at most to render a new page while a full search is running.
Measured on an Apple M1 Max with 32 GB of memory, release build, searching Rust in Action (457 pages, 18 MB) through the reader's worker, three runs. Your results will vary with the document and machine.
Read the code that reads your files.
MIT licensed and built in the open. Every feature lands through a public pull request with tests.
Star on GitHub- Language
- Rust
- Interface
- GPUI, GPU-accelerated and native
- Rendering
- PDFium
- Agents
- Agent Client Protocol and MCP
- Platforms
- macOS on Apple silicon and Intel, Windows x64
- License
- MIT
Start reading. The agent is on its way.
Current builds are not yet signed for every platform. On first launch, macOS may ask you to allow the app in Privacy and Security, and Windows SmartScreen may ask you to confirm. Install steps and checksums.
Does Agentic PDF send my documents anywhere?
No. The reader works entirely on your computer. When the agent panel arrives, your document is only shared with the agent you choose, through that agent's own connection.
Which agents will work?
Agents that speak the Agent Client Protocol, including Claude Code, Codex and Gemini CLI through their ACP adapters. You use your existing subscription.
Is it free?
Yes. Agentic PDF is free and open source under the MIT license.
Can it read scanned PDFs?
Not yet. Pages without a text layer are reported as unreadable. On-device OCR is being considered for the agent release.