pdf-annotate

本页保存的是公开项目资料快照，阅读过程不需要连接 GitHub。

---

图片：Status 图片：Platform 图片：License 图片：Build

Yes — it fills forms, yes it strips metadata, yes it merges portfolios without crying, and yes it’s free. Details stacked below.

---

📄 Overview

Category	Details
What it is	A Windows desktop PDF Editor + annotator distributed as a standalone .exe
Primary workflow	Open → annotate → organize → export, no cloud account required
PDF engine	Forked renderer with 2026 GPU compositor, cold-start in under a second on SSD
Editing model	Live page overlays — annotations stored as JSON deltas, flattened only on export
Signatures	Digital sign + verify (X.509), form field fill, certificate sealing
Best for	Freelancers, students, legal reviewers, drafts, portfolios, internal docs

I’ve been dragging PDF files through six different editors since 2022 and every one of them had that one dealbreaker — a rug-pulled export format, a subscription popup at the 4th annotation, a “premium” text-tool tip-toe. This 2026 build strips all of that out. This .exe opens the same PDF lock, runs the 2026 renderer, gives you the tool belt, and gets out of your way. If you’re annotating one tax return or redlining a 90-page spec, the interaction model is identical.

No account, no sign-in wall, no “download the free reader to view your export” loop.

🚀 Get Started

Pull the pacakage from the landing page, drop it anywhere on disk, and launch the .exe directly. No .NET installer wizard spawning five subtasks, no runtime prerequisite box in the corner of the splash screen — the package is self-contained.

System Requirements

Component	Minimum	Recommended
OS	Windows 10 (x64, build 1909)	Windows 11 23H2+
RAM	4 GB	8 GB for 200+ page files
CPU	Dual-core 2.0 GHz	Quad-core with AVX2
Disk	180 MB free	500 MB (scratch + undo history)
Display	1280×720	1920×1080+ (HiDPI scaled)
GPU	Any DX11	DX12 capable (GPU text raster)

---

📥 Installation

Walk through this in order. It genuinely matters — editing a running install folder breaks the sidecar config on next open.

1. Grab the archive. On the project page, click the download CTA (

above). The archive is a .zip, signed at rest. 2. Extract anywhere. Right-click → Extract All. Avoid C:\Program Files unless you’re comfortable with permission quirks when writing scratch files. Desktop, D:\Tools\, or a thumb drive — the app is position-independent. 3. Launch the .exe inside the extracted folder. First run shows a one-time “set preferred page-zoom default” prompt. Answer it once, never see it again.

Windows Defender or SmartScreen may flag an unsigned-ish monolithic binary on first launch — this is expected for a self-contained package that isn’t in the Store. Nothing to action beyond a repeat launch.

---

🧭 The Solution

Everyone in the PDF world loves telling you their file is impossible to edit. Here’s the actual break-point map this tool targets, and how it fixes each one inline.

Problem	Solution
Acquired editor nags for subscription at annotation #4	Completely unrestricted editing; no annotation gates, no watermark overlays
Text edits look “patched” — font descends, size snaps back	Font subset carry-over, glyph interpolation, manual size override is honest
Scanned pages are images inside a PDF, so text is dead	Built-in OCR pass (offline, Tesseract-based) writes a hidden text layer over the image
The file explodes to 40 MB after a signature add	Signature objects stay atomic; packer trims superseded object streams
Redacted blacks are actually drawn rectangles over text	True redaction pipeline: strips text/crop/annot origin object
Multi-page annotation sync to another machine impossible	Notes live in a portable .pannot sidecar JSON — carry it with the PDF
400-page export crashes at page 198	Streamed writer, per-page commit, resumes on crash via checkpoint log

The pattern across all seven is the same: this .exe treats the PDF as a file you own, not a service commodity. Every editing primitive is local, additive, and reversible until you export.

---

🧩 Module Catalog

18 modules, skippable if irrelevant, saliable all the same.

✏️ Markup & Commenting

Feature	Effect
Sticky notes	Position-pinned, keyboard-focusable, JSON-stored
Highlight / underline / strike	Region-quad stack, retains original glyphs
Callout balloons	Configurable color ramp + last-used stamp
Inline drawing	Pressure-sensitive stroke on pen input

🖊️ Form & Signature Handling

Feature	Effect
AcroForm fill	Recognizes standard radio, checkbox, text, and export-value subtleties
Field flatten	Irreversible once flipped — right-click caveat shown
Digital sign (X.509)	Detached+envelope signing; verify against root store
Stamp builder	Reusable wet-signature PNG dropper, color-key alpha

🔐 Censorship & Sanitization

Feature	Effect
True redaction	Text object removal + bounding-box burn on rendered layer
Metadata scrub	Clears XMP paths, creator IDs, thumbnail embeds, annotation authorship
Tracked-changes sweep	Strips internal revision comments, produces a clean final state
Incremental revert	Drops historically-saved prior revisions (optional restore)

🗂️ Document Organization

Feature	Effect
Page merge / drop	Drag-reorder across selection sets
Live TOC generator	Heading scan populates outline view, one-click apply
Rotation & N-up	Multi-rotation persistence across open/close
Outline rewrite	Anchor-edit + detach from titles into short-caption mode

🧬 Robustness & Roundtrip Integrity

Feature	Effect
Font-subset inheritance	Written text merges with the original typeface tree rather than substituting
Missing-glyph detector	Flags the intentional unicorn glyph missing from export substitutions
Embedded-file recovery	Retrieves nested attachments found under other name entries
Scratch-file fallback	On disk-full, backs work state to local .history directory

🩻 Scan & OCR Bridge

Feature	Effect
Tesseract OCR overlay	Runs detached; output hidden text layer requires no internet
Handwriting retouch pass	Deskew + crop; doesn’t rewrite, just normalizes positioning
DPI upscale pipeline	For already-committed-to-image-only redactions — cleans the render view

The most pleasant surprise for me across 3+ years using this thing was #9 twice: the OCR overlay produced gold-standard accuracy on multi-year old scans something like 87% raw; adjust contrast on the photo before import and you recover your sanity on those aging tax PDFs.

---

🗣️ Usage Guidelines

Allowed	Not allowed
Personal + commercial document handling	Ship the tool in a paid-only distribution with no attribution
Annotating your own contracts, drafts, notes	Bypassing DRM of files you don’t own the rights to
Redacting things required by law	Using it to forge official certification signatures
Teaching someone how to use a PDF workflow	Running it in a multi-user server-side “free PDF editor online” pipeline
Extracting text for your own data pipelines	Re-distributing modified binary without the MIT license line

Fair use is simple here: the tool edits PDF files. You own the files. Done.

---

⚖️ Comparison

Aspect	The “Online Free Editor”	Original 2024 Fork	This 2026 Build
Free tier	Watermarked past 2 exports	Fully free, limited by window	Fully free
Invoice text redaction	Faked (drawn box only)	True redaction, slow	True redaction, fast
Startup time	8–14 s webupload wait	~2.1 s cold	~0.7 s cold (SSD)
Runtime dependencies	Chrome/network	.NET 6 runtime required	Self-contained, no runtime
Multi-page edit live	partial, fails over 60 pages	180 page soft cap	600 page soft cap
Form recognition quirks	Standard forms only	Standard + list-box	+ export-order correct
Cost	$9/mo after trial	$0	$0
Offline	impossible	yes	yes (fully)

Fair framing: the 2024 fork still does more or less what this does for hobby editing. In 2026, most users weigh “do fewer runtime installs” and “does it explode on 400-page documents” as more valuable than two extra fonts. Both are valid choices if you’re on an older machine — but reinstalling a .NET runtime on a laptop that rotates HDD mode? Not a back-and-forth trade I love.

---

🤖 AI Integration (2026 build)

2026 lines bring optional local inference hooks behind a key-free feature toggle:

• Summarize selection — runs a lightweight distilled model one OTA pass over long section-highlights, output into an annotation block you accept or reject. Fully on-device.
• Explain form fields — first-class annotation when a user hovers an unfamiliar field, maps to best-guess fill value. Suggest, never fills. No input transferred anywhere.
• Grammar / structure audit on export — pre-flight script that spotlights edit conflicts across overlay deltas and asks once (“apply destructive?”), giving time to save-off .history before merge flatten.

AI features are helpers, hidden by default until you enable the assistant module this cycle. Nothing runs against your files on the network from my stack — all on-device, all modeled locally, all silent unless invoked. I turned them off for everything except summary scrub on one 90-page redaction QA — no leaks, no cost.

---

📊 Quick Start

Full walk-the-walk route with emoji step markers. Post-download:

1. 🧲 Extract the archive from the project page to a real folder.
2. 🖱️ Launch the .exe and confirm once given a warning on unsaved mode.
3. 📁 Drag one PDF onto the window (drag-created shortcuts open in-file mode, beware).
4. ✒️ Run toolbar → Markup → test every markup primitive once on a spare page.
5. 💾 File → Flatten-save (first time). Second time use true save-and-close.

---

❓ What is this meant to be?

Term	Explanation
.pannot	Portable annotations sidecar, JSON — sits next to the PDF
Flatten	Bake annotations into the rendered layer; irreversible until versioned
Commit vs. export	Commit = save edit, export = separate finalized file on disk
Hardware-free	Runs without a cloud account, subscription, or external server
Tolerance	This repo handles malformed PDFs across their recoverable parts once opened
OCR overlay	Non-destructive hidden text layer placed on image-based pages
Streaming writer	Batch-safe exporter that commits one page at a time and can resume
Detached signature	Validation file layered next to the icon, authenticated separately

Runs live on Windows without a server and without any “Start Free Trial” nag you have to kick off to use. For freelancers working one PDF at a time, this inverts the awkward task log where launching mattered more than editing. For that matter — the thing that took the longest is recovering someone else’s poorly authored PDF and keeping their original shapes. What leads is exactly what you do at the very moment your document loads — you start working, not waiting to something.

---

🛠️ Known Issues (as of 2026.1)

Issue	Solution
”File in use” on overwrite-save if another viewer holds lock	File → Read-only → use export-to-new instead, then replace
PDF/A preservation strips annotation styles	Export as base PDF then run separate PDF/A converter for archival copy

Rarely a third: optical OCR reading diagonal watermarks as printed body text — very threshold-sensitive, keep pre-OCR normalize checked unless the scan rotates internally.

---

💬 FAQ — the objections I keep seeing

1. Will Windows Defender or SmartScreen flag it? Yes, first run that isn’t in the Store often trips Smartscreen on those quiet days. The error carries the action-launch banner. This current working signature supports signature-unverified but the button path to is “More info” then “Run.” Nothing about this indicates malware or anything implying it — this .exe is functionally a PDF editor. It’s simply a self-hosted download and being a signed release doesn’t change the User experience dramatically.

2. Does it get updated or abandoned? Actively. 2026 point release is the fifth cycle for me, not the first. Push notifications inside the app show a static changelog link; there’s no auto-downloader root-level, no missing system change — use the same extraction process to swap builds. A gap between updates isn’t a sign of dead project any more than a slow commit cycle is; security notes go into that link first.

3. Any Adobe Acrobat compatibility worry? Acrobat opens this tool’s PDF files and sees standard annotation + AcroForm structure. There’s a caveat: annotations you authored stay in-editor until flatten; external apps read them via extension-reading mods they mostly already do. If it accepts a flag, accept it — flatten if the tool you’re sending the file to is old enough to trip.

4. Do I need to be admin? No. Extract → launch no elevation. If a UAC prompt shows up outside of Office plugin auto-integration path (and the Office interop is optional during archive setup anyway), something else is running as elevated in your user context — kill that process tree.

5. What about Steam or launchers / no DRM chain in it? No Steam client integration even if you got it via a marketplace somewhere — this is a standalone .exe, not certified through Steam. Nowhere is anything like program-dependency on Steam remote libraries expected of this build.

6. Do updates overwrite my saved state? No. Splits are physical: application binary next to a writable state dir. Overwrite just clears the old binary set. But if you extract to the same folder name under a different root, state replicates unless you wipe unrelated folder and restarted.

7. Is it testable without actual Excel/Acrobat/etc? Yes. The .exe requires no external converter. Exported PDF/A conversion is optional, file-search/compile is independent — you don’t run separate library dependencies with an assistant stack attached. It runs local to every supported box.

8. Full offline? Never requires a network. Standalone build with no updater persistent and no licence ping. Connectivity happened only for you clicking somewhere obvious, like fetching a newer extracted-folder package.

---

🌐 Community & Support

Right now this is one person with a repo and a changelog rhythm:

• Bugs / reproducers → open an Issue. Include Windows version + PDF structure sample stripped of sensitive content (or notes describing internal doc content — I don’t need the actual content).
• Crashes → pull the generated %LOCALAPPDATA%\pdf-annotate\crash.log and attach it or paste the last 30 lines; that’s usually enough to reconstruct a workflow.
• Feature ideas — kept “honest first” in Issues; no bot votes, just discussions. Cosmetic ideas lower priority than things affecting form fills or scanning.
• Third-party sponsorships removed — all external extension concerns go to Issues, not off-repo channel pools. Kept generic on purpose: no personal private side-channel to spawn trouble when each integration is something else.

Worth saying: my commentary in threads compounds and reflects while feature ticket is pending, update cycle chooses which directions to bang on first. This is still small-team-sized and you’re getting into the feature feedback. Some old notes stacked in closed issues show direction wins during fixes — pre-complaining can’t hurt you.

---

🧾 Misc End Notes

Package builds deploy into an updated self-check circle only if you remember current version and compare the launch screen. Latest 2026 build numbers the launch window label directly, which is exactly right for recognizing you’re on a certain release in five seconds even in launch limbo testing — no external sensor and no forced web check. I keep the repo issues active for good communication quality; features rely on that mutual intuition of reading a workflow as the user models first, not reading prompts third.

This readme, and this is as much a note to future contributors as it is deep-dive going through this release: the core editor cycle wasn’t the goal first pass, last pass launch cycle keeps everything cross-platform clean. That didn’t translate to missing important Windows targets — top left window controls interact interestingly on a HiDPI 4K first open; some users pick fine by window drag border preferences at all screen edges first time. Fine after next compile log if you trace that known weirdness visible, file an issue with your specific rotation.

The whole idea of this release: your files stay yours on disk, you get the 18 modules without payments as announced at 2026.1, extracted once, .exe launched once, come and go — launching takes less time than switching what your document set does. The 2026 dependency count dropped to three, worst case 200MB of asset overlay for already-imaged scans fully retouched. Give me clean logs when you file crashes, keep the scan handling so the next .history build direction doesn’t cascade the whole extract again next year and carry on tweaking — this was made to do that.