pdfeverything

✨ Why PDFeverything?

You've been there: a Word report here, a PDF scan there, some photos of the whiteboard, an Excel chart… and now someone wants "one combined PDF please, by 5pm". 😤

PDFeverything is built for exactly this moment.

Drag everything in — any combination of PDFs, Word documents, PowerPoint decks, Excel spreadsheets, PNGs, JPEGs, text files — and it merges them into one unified PDF, in the order you decide, with progress you can watch.

🔗 Latest Release →

🪟 The Windows download is an installer — bilingual wizard, licence page, Start Menu entry, optional desktop shortcut and a standard uninstaller. No Python required.

Related MCP server: gurupdf-mcp

🆕 What's New in v1.9.0

A QA-led hardening round: 19 defects fixed, hot paths made several times faster.

Every item below came out of an adversarial test pass (malformed input, boundary values, hostile page ranges, cross-channel consistency, memory accounting) plus an independent audit of the reader. Each fix has a regression test that failed before it.

• 🧠 The reader's memory ceiling is real now. The 400 MB pixmap cache stopped evicting the moment it met a 100%-zoom page — the exact entries a freshly opened document is full of — so it grew without limit (measured 638 MB against a 400 MB cap). Eviction is now two-tier: disposable zoomed renders first, then off-screen bases, never the page you are looking at. It also charged the device pixel ratio twice, over-counting 4× on Retina and making the cap behave like 100 MB.

• 🖱️ Grid clicks hit the page you clicked. Grid cells were hit-tested with label-local mouse coordinates against a container-space grid, so every thumbnail resolved to page 1 — "click page 5, press Delete" deleted page 1. The left edge of any thumbnail produced the phantom index -1, which could duplicate the last page during a drag-sort.

• 🛡️ Drag-sort no longer aborts the app. Reordering rebuilt the document and left the reader holding a closed handle; the next line raised inside a Qt slot, which terminates the process and lost unsaved edits with no prompt.

• ⚡ Zoom and resize are ~10× faster on long documents. Layout read doc[i].rect for every page on every scroll-stop, zoom tick and resize — 7.5 ms of a 9.5 ms pass at 1000 pages, i.e. most of a 16.7 ms frame budget per pinch tick. Page geometry is now cached: 9.5 ms → 0.94 ms per layout, 11.9 ms → 1.55 ms per pinch tick. A 1000-page PDF still opens in 30 ms using 10 MB.

• 🪶 Edits stay lazy. Deleting a page re-rendered the entire document at 100% on the GUI thread (189 MB for 100 pages, ~1.9 GB at 1000). Only the visible rows are materialised now — the windowed renderer finally holds for edits too.

• ↩️ Undo history actually works from the CLI. The journal was keyed on (path, size, mtime), so a command's own output never matched its input and page-undo always answered "nothing to undo". Commands now edit in place when -o is omitted, and the journal identity is recorded so it survives the rewrite. Journal snapshots are also byte-budgeted and stale ones are swept — it used to keep 50 full copies of the document in RAM and on disk forever.

• 🔐 Password-protected PDFs explain themselves. Every operation used to leak a library internal ("File has not been decrypted", "document closed or encrypted", "PasswordError"). All 16 entry points now return one actionable message.

• 🧩 One page-range grammar for all three channels. The MCP server had its own splitter that rejected what the CLI accepted (1–5, 1，3, 3-1, 1..5) and crashed on others. GUI, CLI and MCP now share core.utils.parse_page_ranges, and truncated specs like 1- are refused instead of silently meaning 1.

• 📐 Text→PDF wrapping measures instead of guessing. A fixed 6.5 pt/char estimate left Latin text 85 pt short of the margin and pushed CJK 54 pt past the page edge. Chunks are now measured with the real font.

• 🧹 CLI stops lying. An unknown command silently opened the GUI with exit code 0 (a failed scripted call looked successful); it now prints a suggestion and exits 2.

• 🎯 Per-document state resets on open. Search hits, highlights and edit mode used to survive a file switch: a 100-page document's results stayed in the sidebar over a 3-page one, the page label could read "61 / 3", and old hit rectangles were baked into the new document's pages.

• 🖼️ Fit-mode renders follow a resize. The fit cache key ignored the viewport, so after resizing, the label kept the pre-resize pixmap (424×600 inside a 809×1145 frame) and never filled the window again.

• 🧪 A hostile-file pass. 0-page, 0-byte, truncated, fake-header and password-protected files are now opened and edited without a crash; a failed open reports what went wrong instead of leaving a blank dead-end.

• 🧰 Plus: reading position is saved on ⌘Q (not only on ✕), a force-terminated worker releases the UI, the tooltip uses one timer instead of one per hover, JPG export honours the filename you chose, and a stray print() that corrupted the MCP JSON-RPC stream is gone.

• 🧪 New test gates: tests/qa_adversarial.py (170 probes: robustness, accuracy, cross-channel consistency, performance and temp-file hygiene) and tests/qa_reader_defects.py (37 reader regressions, one per defect above). Full set: **93 core + 70 GUI + 38 reader + 85 reader-comprehensive + 37 reader-defects

  • 170 adversarial = 493 checks**.

🎯 The Killer Feature: Mixed-File Merge

📄 report.docx   (2 pages)
📊 chart.xlsx    (1 page)
🖼️ photo.jpg    (1 page)
📄 appendix.pdf  (5 pages)

         🪄  one click  🪄
              ↓
    ┌─────────────────────┐
    │   unified.pdf       │
    │   9 pages, in order │
    └─────────────────────┘

Every file goes through its own converter (AppleScript → Office on macOS, or pure-Python fallback), then everything gets stitched together. If one file fails, the rest still go through — you get a summary of what worked and what didn't.

📖 Built-in PDF Reader

A high-performance PDF reader with vector-grade rendering — rivaling Acrobat and WPS.

🔧 Everything It Can Do

📥 Supported Inputs

🖥️ The GUI

┌──────────────────────────────────────────────┐
│  📄 doc.pdf        [🔼] [🔽] [✖]            │
│  📊 data.xlsx      [🔼] [🔽] [✖]            │
│  🖼️ photo.jpg      [🔼] [🔽] [✖]            │
│  📝 report.docx    [🔼] [🔽] [✖]            │
│                                              │
│  ── drag & drop files here ──               │
│                                              │
│  [🔀 Merge All →]       [✂️ Split...]        │
│  [🗜️ Compress...]       [💧 Watermark...]    │
│  [🔒 Encrypt...]        [🔄 Rotate...]       │
│                                              │
│  ████████████░░░░░░  78%                     │
│  Converting: report.docx (3/5)...            │
└──────────────────────────────────────────────┘

• 🖱️ Drag & drop files from Finder / Explorer

• 🔄 Reorder with arrow buttons or by dragging inside the list

• ⚡ Multi-threaded — never freezes, always shows progress

• 🧠 Smart buttons — the UI adapts to what's in your file list

• 🌐 Bilingual UI — switch between Chinese and English (Settings > Language)

🤖 AI Agent Integration (MCP Server)

PDFeverything comes with a built-in Model Context Protocol (MCP) server. Any AI agent (Claude Desktop, Claude Code, Cursor, etc.) can discover all 29 PDF tools and call them directly — no Python, no install, just the app file.

How it works

The same .exe / .app binary supports three modes:

Setup — Claude Desktop

Add to ~/.claude/claude_desktop_config.json:

{
  "mcpServers": {
    "pdfeverything": {
      "command": "/Applications/PDFeverything.app/Contents/MacOS/PDFeverything",
      "args": ["--mcp"]
    }
  }
}

Windows:

{
  "mcpServers": {
    "pdfeverything": {
      "command": "C:\\Program Files\\PDFeverything\\PDFeverything.exe",
      "args": ["--mcp"]
    }
  }
}

Setup — Claude Code

Add to .claude/settings.json in your project:

{
  "mcpServers": {
    "pdfeverything": {
      "type": "stdio",
      "command": "/Applications/PDFeverything.app/Contents/MacOS/PDFeverything",
      "args": ["--mcp"]
    }
  }
}

What the AI sees (29 tools)

Once connected, the agent automatically discovers these tools — no manual instruction needed:

Direct CLI mode (no MCP needed)

AI agents can also call the binary directly:

# macOS
/Applications/PDFeverything.app/Contents/MacOS/PDFeverything merge -i a.pdf b.pdf -o out.pdf
/Applications/PDFeverything.app/Contents/MacOS/PDFeverything info -i doc.pdf
/Applications/PDFeverything.app/Contents/MacOS/PDFeverything -h

# Windows
PDFeverything.exe merge -i a.pdf b.pdf -o out.pdf
PDFeverything.exe info -i doc.pdf
PDFeverything.exe -h

💡 The .app and .exe are the SAME single binary. Give it CLI args → headless mode. Give it --mcp → MCP server. No args → GUI. One file, three personalities.

🚀 Quick Start (for Developers)

# 1. Install dependencies
pip install PyQt6 PyMuPDF pypdf pikepdf pillow python-docx python-pptx openpyxl

# 2. Launch GUI
python main.py

# 3. Or use the CLI
python pdf_tool.py merge -i a.pdf b.pdf -o merged.pdf
python pdf_tool.py info -i document.pdf

# 4. Or start the MCP server
python mcp/server.py

Build from Source

Windows (installer + portable exe, needs Inno Setup 6):

python build_windows.py
# → PDFeverything_Setup_v1.9.0.exe   installer (copied to the project root)
# → dist/PDFeverything.exe           portable payload, no install

macOS (app bundle):

pyinstaller PDFeverything.spec --noconfirm --clean
# → dist/PDFeverything.app

🧪 Testing

Three gates run on every release, plus two adversarial suites added in v1.9.0:

.venv/bin/python tests/test_core.py                            # core / CLI / MCP / i18n   93
QT_QPA_PLATFORM=offscreen .venv/bin/python tests/test_gui.py    # workers / batch / dialogs 70
QT_QPA_PLATFORM=offscreen .venv/bin/python tests/qa_reader.py   # reader QA                 38
QT_QPA_PLATFORM=offscreen .venv/bin/python tests/test_reader.py # reader comprehensive      85

# adversarial suites — these attack the software, they do not just re-check it
.venv/bin/python tests/qa_adversarial.py                        # robustness / accuracy   170
QT_QPA_PLATFORM=offscreen .venv/bin/python tests/qa_reader_defects.py  # reader defects    37

tests/qa_adversarial.py covers malformed and hostile input (0-byte, truncated, fake-header, encrypted, 0-page PDFs), page-range boundary values, cross-channel consistency between GUI/CLI/MCP, pixel-level accuracy oracles, memory and file-descriptor ceilings, and temp-file hygiene. Every probe prints the concrete evidence when it fails.

tests/qa_reader_defects.py holds one regression per reader defect fixed in v1.9.0; each of them failed before its fix.

Total: 493 checks.

🧱 Tech Stack

📄 License

MIT — do whatever you want with it. LICENSE

🎯 核心功能：混合文件合并

📄 报告.docx     (2 页)
📊 图表.xlsx     (1 页)
🖼️ 照片.jpg     (1 页)
📄 附录.pdf      (5 页)

         🪄  一键合并  🪄
              ↓
    ┌─────────────────────┐
    │   统一输出.pdf       │
    │   9 页，顺序不变     │
    └─────────────────────┘

把不同类型的文件拖进来，每个文件经过专用转换器处理，然后按顺序合并。某个文件转换失败了也不影响其他的——最后会给你一份汇总报告。

📖 内置 PDF 阅读器

高性能 PDF 阅读器，矢量级渲染画质——媲美 Acrobat 和 WPS。

🔧 全部功能

🆕 v1.9.0 新功能

以 QA 为主导的加固迭代：修复 19 个缺陷，热点路径提速近 10 倍。

下面每一条都来自对抗性测试（畸形输入、边界值、恶意页码范围、三通道一致性、 内存计量）以及一次独立的阅读器审计。每个修复都配有"修复前必然失败"的回归测试。

• 🧠 阅读器的内存上限真正生效了 — 原先淘汰循环一遇到 100% 缩放的页面就 break， 而刚打开的文档里全是这种页面，于是缓存无上限增长（实测 638 MB 对 400 MB 上限）。 现在分两级淘汰：先丢可再生的缩放/缩略图渲染，再丢屏幕外的 base，绝不丢你正在看的那一页。 同时修正了重复计算 devicePixelRatio 的问题（Retina 上 4 倍高估，使 400 MB 实际只有 100 MB）。

• 🖱️ 网格视图点哪页选哪页 — 网格单元用标签局部坐标去做容器坐标的命中测试， 导致每个缩略图都解析成第 1 页：点第 5 页再按删除，删掉的是第 1 页。缩略图左边缘还会产生 幽灵索引 -1，拖拽排序时可能复制最后一页。

• 🛡️ 拖拽排序不再让程序崩溃 — 重排会重建文档，阅读器却仍持有已关闭的句柄， 下一行在 Qt 槽函数里抛异常 → 进程直接 abort，未保存的修改无提示丢失。

• ⚡ 长文档缩放/缩放窗口快约 10 倍 — 布局在每次滚动停止、缩放、改变窗口时都要为 每一页读 doc[i].rect（1000 页时占 9.5ms 中的 7.5ms，等于一次捏合缩放吃掉大半个帧预算）。 页码几何现已缓存：布局 9.5ms → 0.94ms，捏合一次 11.9ms → 1.55ms。 1000 页 PDF 打开仍是 30ms、占用 10MB。

• 🪶 编辑后依然惰性渲染 — 删除一页曾在 GUI 线程上把整本文档按 100% 全部重渲 （100 页 189MB，1000 页约 1.9GB）。现在只渲染可见行。

• ↩️ CLI 的撤销历史真的能用了 — 日志键基于 (路径, 大小, mtime)， 导致命令自己的输出永远匹配不上输入，page-undo 总是回答"没有可撤销的操作"。 现在省略 -o 即就地修改，且日志身份被持久记录，改写后依然能找回历史。 日志快照同时按字节预算裁剪，陈旧日志会被清理（原先会在内存和磁盘各留 50 份完整副本）。

• 🔐 加密 PDF 给出可操作的提示 — 原先 16 个入口各自泄漏底层库的内部错误 （"File has not been decrypted"、"document closed or encrypted"、"PasswordError"）。

• 🧩 三通道统一页码语法 — MCP 曾有一套自己的解析器，CLI 能识别的 1–5、1，3、 3-1、1..5 它会报错，另一些则直接崩溃。现在 GUI/CLI/MCP 共用 core.utils.parse_page_ranges，且 1- 这类残缺写法会被明确拒绝而不是悄悄当成 1。

• 📐 文本转 PDF 按实际宽度折行 — 原先按 6.5pt/字符估算：拉丁文比右边距短 85pt， 中文则超出页面 54pt。现在用真实字体测量。

• 🧹 CLI 不再"假装成功" — 未知命令原先会静默打开 GUI 并返回 0（脚本调用失败看起来像成功）， 现在会给出相近命令建议并以退出码 2 结束。

• 🎯 切换文件时重置每文档状态 — 搜索结果、高亮、编辑模式原先会跨文件残留： 100 页文档的结果留在 3 页文档的侧栏里，页码可能显示 "61 / 3"， 旧文档的命中矩形还会被画进新文档的位图。

• 🖼️ 适应窗口模式跟随窗口尺寸 — fit 缓存键没有包含视口尺寸，改变窗口后 标签仍拿着改变前的位图（424×600 塞在 809×1145 的框里），永远不会铺满。

• 🧪 恶意文件专项测试 — 0 页、0 字节、截断、伪造文件头、加密 PDF 现在都能 正常打开/编辑而不崩溃；打开失败会说明原因而不是留下一片空白。

• 🧰 其他：⌘Q 退出也会记住阅读位置（原先只有点 ✕ 才记）、强制终止的 worker 会释放界面、 工具提示改为单个复用定时器、JPG 导出遵循你选的文件名、 以及一处会污染 MCP JSON-RPC 流的 print() 已移除。

• 🧪 新增测试门禁：tests/qa_adversarial.py（170 项：鲁棒性/准确性/三通道一致性/ 性能/临时文件卫生）与 tests/qa_reader_defects.py（37 项阅读器回归，每个缺陷一项）。 全套件：93 核心 + 70 GUI + 38 阅读器 + 85 阅读器综合 + 37 阅读器缺陷 + 170 对抗 = 493 项。

🖥️ 界面预览

┌──────────────────────────────────────────────┐
│  📄 doc.pdf        [🔼] [🔽] [✖]            │
│  📊 data.xlsx      [🔼] [🔽] [✖]            │
│  🖼️ photo.jpg      [🔼] [🔽] [✖]            │
│  📝 report.docx    [🔼] [🔽] [✖]            │
│                                              │
│  ── 拖放文件到此处 ──                        │
│                                              │
│  [🔀 合并为统一 PDF]   [✂️ 拆分...]          │
│  [🗜️ 压缩...]        [💧 水印...]          │
│  [🔒 加密...]        [🔄 旋转...]           │
│                                              │
│  ████████████░░░░░░  78%                     │
│  转换中: 报告.docx (3/5)...                  │
└──────────────────────────────────────────────┘

• 🖱️ 从资源管理器/Finder 拖放文件

• 🔄 用按钮或拖动调整顺序

• ⚡ 多线程处理 — 界面永不卡顿，实时进度条

• 🧠 智能按钮 — 根据文件列表内容自动变化

🤖 AI Agent 集成（MCP 服务器）

PDFeverything 内置了 Model Context Protocol (MCP) 服务器。任何 AI Agent（Claude Desktop、Claude Code、Cursor 等）都能自动发现全部 29 个 PDF 工具并直接调用——无需安装 Python、无需额外依赖，只要有这个 app 文件就行。

同一个文件，三种模式

配置 Claude Desktop

在 ~/.claude/claude_desktop_config.json 中添加：

{
  "mcpServers": {
    "pdfeverything": {
      "command": "/Applications/PDFeverything.app/Contents/MacOS/PDFeverything",
      "args": ["--mcp"]
    }
  }
}

Windows：

{
  "mcpServers": {
    "pdfeverything": {
      "command": "C:\\Program Files\\PDFeverything\\PDFeverything.exe",
      "args": ["--mcp"]
    }
  }
}

配置 Claude Code

在项目下的 .claude/settings.json 中添加：

{
  "mcpServers": {
    "pdfeverything": {
      "type": "stdio",
      "command": "/Applications/PDFeverything.app/Contents/MacOS/PDFeverything",
      "args": ["--mcp"]
    }
  }
}

AI 能看到的 29 个工具

连接后 Agent 会自动发现这些工具——无需手动教它：

直接 CLI 调用（无需 MCP）

AI Agent 也可以直接调用二进制：

# macOS
/Applications/PDFeverything.app/Contents/MacOS/PDFeverything merge -i a.pdf b.pdf -o out.pdf
/Applications/PDFeverything.app/Contents/MacOS/PDFeverything info -i doc.pdf
/Applications/PDFeverything.app/Contents/MacOS/PDFeverything -h

# Windows
PDFeverything.exe merge -i a.pdf b.pdf -o out.pdf
PDFeverything.exe info -i doc.pdf
PDFeverything.exe -h

💡 .app 和 .exe 就是同一个二进制文件。 给参数 → 命令行模式。给 --mcp → MCP 服务器。不给参数 → GUI。一个文件，三种用法。

🚀 开发者快速开始

pip install PyQt6 PyMuPDF pypdf pikepdf pillow python-docx python-pptx openpyxl
python main.py            # 启动 GUI
python pdf_tool.py info -i document.pdf   # CLI 模式

📦 从源码构建

Windows（安装包 + 便携版 exe，需要 Inno Setup 6）:

python build_windows.py
# → PDFeverything_Setup_v1.9.0.exe   安装包（同时复制到项目根目录）
# → dist/PDFeverything.exe           便携版载荷，免安装

macOS（App Bundle）:

pyinstaller PDFeverything.spec --noconfirm --clean
# → dist/PDFeverything.app

🧪 测试

发布前跑三个门禁套件，v1.9.0 起再加两个对抗性套件：

.venv/bin/python tests/test_core.py                            # 核心 / CLI / MCP / i18n   93
QT_QPA_PLATFORM=offscreen .venv/bin/python tests/test_gui.py    # Worker / 批量 / 对话框   70
QT_QPA_PLATFORM=offscreen .venv/bin/python tests/qa_reader.py   # 阅读器 QA                38
QT_QPA_PLATFORM=offscreen .venv/bin/python tests/test_reader.py # 阅读器综合测试            85

# 对抗性套件 —— 它们主动攻击软件，而不只是复检既有行为
.venv/bin/python tests/qa_adversarial.py                        # 鲁棒性 / 准确性         170
QT_QPA_PLATFORM=offscreen .venv/bin/python tests/qa_reader_defects.py  # 阅读器缺陷回归    37

tests/qa_adversarial.py 覆盖畸形与恶意输入（0 字节、截断、伪造文件头、加密、0 页 PDF）、 页码范围边界值、GUI/CLI/MCP 三通道一致性、像素级准确性判据、 内存与文件描述符上限、临时文件卫生。每一条失败时都会打印具体证据。

tests/qa_reader_defects.py 中每个 v1.9.0 修复的阅读器缺陷都对应一项回归， 它们在修复前必然失败。

合计 493 项。

🧱 技术栈

📄 许可证

MIT — 随便用。 LICENSE