pdf-insight-mcp 0.2.1

MCP server for extracting text, images, tables, links, annotations, and metadata from PDF files

pdf-reader-mcp

一个功能丰富的 PDF 阅读 MCP 服务器，让 LLM（大语言模型）客户端能够读取和分析 PDF 文件。
A feature-rich MCP server for reading and analyzing PDF files with LLM clients.

功能特性 / Features

工具 / Tool	中文说明	English
get_pdf_info	读取文档元数据、页数、大小和加密状态	Read document metadata, page count, size, and encryption status
read_pdf_as_text	提取指定页面文本内容	Extract text content from selected pages
read_pdf_as_images	将指定页面渲染为 base64 图片	Render selected pages as base64-encoded images
get_pdf_outline	读取书签与目录结构	Read bookmarks and outline structure
search_pdf_text	按页返回搜索结果和上下文	Search text with per-page context
extract_pdf_tables	提取可识别的表格结构	Extract structured tables when detectable
extract_pdf_images	提取 PDF 内嵌图片	Extract embedded images from the PDF
get_pdf_page_info	查看单页尺寸、文本、图片和链接信息	Inspect a page's dimensions, text, images, and links
extract_pdf_links	提取外部链接和内部跳转	Extract external URLs and internal page jumps
get_pdf_annotations	读取批注、高亮与注释信息	Read comments, highlights, and annotation data
get_pdf_text_stats	统计文本、行数、段落数和扫描版概率	Compute text, line, paragraph, and scan-likelihood stats
compare_pdf_pages	比较两个页面的文本相似度	Compare text similarity between two pages

为什么做这个项目 / Why this project

很多 LLM 工作流不仅需要纯文本提取，还需要目录、表格、图片、注释、链接等结构化信息。
Many LLM workflows need more than raw text extraction. They also need structure, tables, images, annotations, and links.

这个服务提供统一的 MCP 接口，用于： This server provides a unified MCP interface for:

• 文本型 PDF / text-heavy PDFs
• 扫描版或版式敏感 PDF / scanned or layout-sensitive PDFs
• 表格与图片提取 / table and image extraction
• 元数据与结构分析 / metadata and structure inspection
• 批注与链接分析 / annotation and link analysis

安装 / Installation

前置要求 / Prerequisites

• Python 3.10+
• uv 或其他 Python 环境管理工具 / uv or another Python environment manager

安装 uv / Install uv:

curl -LsSf https://astral.sh/uv/install.sh | sh

Windows PowerShell:

irm https://astral.sh/uv/install.ps1 | iex

从 PyPI 安装 / Install from PyPI

发布后可直接通过 uvx 运行： After the package is published, you can run it directly with uvx:

uvx pdf-insight-mcp

也可以先安装再运行： You can also install first, then run:

python -m pip install pdf-insight-mcp
pdf-reader-mcp

本地开发安装 / Local development setup

uv sync

运行服务 / Run the server

uv run pdf-reader-mcp

在 MCP 客户端中配置 / Configure in an MCP client

PyPI 安装方式示例 / Example config using the published PyPI package:

{
  "mcpServers": {
    "pdf-reader": {
      "command": "uvx",
      "args": ["pdf-insight-mcp"]
    }
  }
}

本地仓库开发配置示例 / Example configuration for a local checkout:

{
  "mcpServers": {
    "pdf-reader": {
      "command": "uv",
      "args": [
        "--directory",
        "/absolute/path/to/pdf-reader-mcp",
        "run",
        "pdf-reader-mcp"
      ]
    }
  }
}

将 /absolute/path/to/pdf-reader-mcp 替换为你的本地仓库路径。
Replace /absolute/path/to/pdf-reader-mcp with your local repository path.

发布 / Release

推荐发布路径： Recommended release path:

1. 发布 Python 包到 PyPI / Publish the Python package to PyPI
2. 发布 server.json 到官方 MCP Registry / Publish server.json to the official MCP Registry

建议使用 GitHub Actions + PyPI Trusted Publishing（OIDC）+ MCP Registry GitHub OIDC。 The recommended automation is GitHub Actions + PyPI Trusted Publishing (OIDC) + MCP Registry GitHub OIDC.

典型发布流程： Typical release flow:

# 1. 修改版本号（pyproject.toml 和 server.json）
# 2. 提交改动
git commit -am "Release v0.2.0"

# 3. 打 tag
git tag v0.2.0

# 4. 推送分支和 tag
git push origin main --tags

工作流会在 v* tag 上： The release workflow will, on v* tags:

• 运行测试 / run tests
• 构建 sdist 和 wheel / build sdist and wheel
• 做 twine check / run twine check
• 发布到 PyPI / publish to PyPI
• 发布到 MCP Registry / publish to the MCP Registry

响应大小与大 PDF 注意事项 / Response size and large-PDF notes

• read_pdf_as_images 返回的是 base64 图片，响应体积会迅速变大。
  read_pdf_as_images returns base64 image payloads, which can grow very quickly.
• 图片渲染仍然限制为最多 20 页。
  Image rendering is still limited to 20 pages per call.
• read_pdf_as_text 现在默认限制为最多 50 页、最多 200000 字符，超限会截断并附带 warning。
  read_pdf_as_text now defaults to at most 50 pages and 200000 characters, and truncates with a warning when needed.
• read_pdf_as_images 现在默认限制总返回负载约 20MB，超限会提前停止并附带 warning。
  read_pdf_as_images now defaults to an overall payload cap of about 20MB and stops early with a warning.
• 对扫描版 PDF，建议优先按小页范围调用，并降低 dpi、使用 jpeg、降低 quality。
  For scanned PDFs, prefer smaller page ranges, lower dpi, jpeg, and lower quality.

开发 / Development

安装开发依赖 / Install dev dependencies:

uv sync --extra dev

运行测试 / Run tests:

uv run pytest

技术栈 / Tech stack

• Python 3.10+
• MCP Python SDK
• PyMuPDF

License

MIT