ofdkit

轻量的 Python 工具，用来读取 OFD 文件，提取里面的文字和发票字段，并能接入 AI 工具。OFD 是中国的国家标准版式文档格式（GB/T 33190），常见于电子发票和公文。

A lightweight Python toolkit for reading OFD documents (GB/T 33190, the Chinese fixed-layout format used for e-invoices and official documents): extract text and invoice fields, and use them from AI tools.

状态：0.3 版，接口以后可能还会调整。更新记录见 CHANGELOG。 Status: 0.3, the API may still change. See the CHANGELOG.

能做什么 / What it does

• 看发票：一条命令显示发票号码、日期、购买方、销售方、金额和明细。 Show an invoice: number, date, buyer, seller, amounts and line items, with one command.
• 把一个文件夹里的发票汇总到 Excel，也能直接读税务平台批量下载的 .zip 压缩包。发票号码保持原样（不会变成 2.4E+19），金额可以直接求和，重复的发票会标出来。 Summarise a folder of invoices, or a .zip archive downloaded from the tax service, into an Excel workbook: invoice numbers stay intact, amounts are numbers, duplicates are flagged.
• 报销帮手：检查发票抬头（购买方名称、税号）对不对；按发票内容给文件改名，可以预览、可以改回去。 For reimbursement: check that invoices are made out to the right buyer, and rename invoice files after their contents, with a preview and an undo.
• 按阅读顺序提取文字，包括模板和复合对象里的文字。 Text in reading order, including text in templates and composite objects.
• 给程序和 AI 用：发票字段输出 JSON，提供 MCP 服务器、Agent Skill 和 MarkItDown 插件。 For programs and AI: invoice fields as JSON, an MCP server, an Agent Skill and a MarkItDown plugin.
• 只用 Python 标准库，没有任何第三方依赖；能读很多不规范的文件（路径写法不对、编码不对、XML 前缀错误等）。 Standard library only, no dependencies; reads many non-conforming files (bad paths, encodings, XML prefixes).

快速上手 / Quick start

1. 安装 Python 3.10 或更新版本。Windows 上安装时请勾选 "Add python.exe to PATH"。 Install Python 3.10 or newer.

2. 打开命令行（Windows 在开始菜单搜索"终端"或"cmd"，Mac 打开"终端"），安装 ofdkit： Open a terminal and install ofdkit:

   pip install ofdkit
   

3. 看一张发票（把文件拖进命令行窗口就能填上路径）： Show an invoice (drag the file into the terminal window to type its path):

   ofdkit 发票.ofd
   

   电子发票（普通发票）
   
   发票号码  00000000000000000001
   开票日期  2026-01-02
   购买方    测试单位甲
             税号 TESTBUYER0001
   销售方    测试单位乙
             税号 TESTSELLER0002
   金额      100.00
   税额      6.00
   价税合计  106.00
   明细      *测试*服务项目  100.00  6%
   开票人    测试人员
   

   （数据是虚构的 / fictional data）

4. 看一个文件夹或 .zip 压缩包里的所有发票和合计，然后导出到 Excel： List every invoice in a folder or .zip archive with totals, then export to Excel:

   ofdkit 发票文件夹
   ofdkit batch 发票文件夹 -o 发票汇总.xlsx
   

   Excel 文件有两张表："发票"每张发票一行，最后是合计；"明细"每个商品或服务一行。重复的发票标黄，读不出来的文件标红并写明原因。合计用的是 SUBTOTAL，筛选以后只算看得见的行。 The workbook has two sheets: 发票 (one row per invoice, with totals; duplicates in yellow, unreadable files in red with the reason) and 明细 (one row per line item). Totals use SUBTOTAL, so they follow filters.

5. 检查发票抬头：加上 --buyer，购买方名称或税号不是你公司的发票会标出来（Excel 里标橙色）。名称和税号都要查，就写两次： Check who the invoices are made out to: with --buyer, invoices for another buyer name or tax ID are flagged (orange in Excel). Repeat it to check both:

   ofdkit 发票文件夹 --buyer 某某有限公司 --buyer 91310000XXXXXXXXXX
   

6. 按发票内容给文件改名，例如 dzfp_2411…_20240528.ofd 改成 2024-05-28_某某公司_4449.00_2411….ofd。不加 --apply 只是预览；同名的 PDF、XML 会一起改；改名记录可以用 --undo 改回去： Rename invoice files after their contents. Without --apply it only previews; PDF and XML copies with the same name are renamed too; --undo restores the old names from the log:

   ofdkit rename 发票文件夹            # 预览 / preview
   ofdkit rename 发票文件夹 --apply    # 改名 / rename
   ofdkit rename --undo 发票文件夹/改名记录_20261028_093000.csv
   

   名字的格式可以改，例如 --pattern "{开票日期}_{购买方}_{商品}"。可以用的字段：开票日期、发票号码、发票代码、发票种类、销售方、销售方税号、购买方、购买方税号、金额、税额、价税合计、商品、原文件名。 The name format can be changed with --pattern; fields are listed by ofdkit rename --help (English names such as {date}, {seller} and {total} work too).

命令行 / Command line

ofdkit 文件.ofd                              # 看发票，不是发票就显示文字 / show an invoice, or the text
ofdkit 文件夹或压缩包.zip                     # 发票列表和合计 / list invoices with totals
ofdkit batch 文件夹 -o 发票汇总.xlsx          # 导出 Excel / export to Excel
ofdkit batch 文件夹 -o 发票汇总.csv           # 导出 CSV（给程序用）/ export CSV for programs
ofdkit 文件夹 --buyer 公司名或税号             # 检查发票抬头 / check the buyer
ofdkit rename 文件夹                         # 按内容改名（先预览）/ rename after contents (preview first)
ofdkit text 文件.ofd                         # 提取文字 / print text
ofdkit invoice 发票.ofd                      # 发票字段（JSON）/ invoice fields as JSON
ofdkit info 文件.ofd                         # 页数、作者等信息 / pages and metadata
ofdkit describe 文件.ofd -o 结构.json         # 打码的结构，用于报告问题 / masked structure for bug reports

show、batch 和 rename 加 -r 会连子文件夹一起找。ofdkit 文件.ofd 是 ofdkit show 文件.ofd 的简写。 Add -r to show, batch and rename to include subfolders. ofdkit PATH is short for ofdkit show PATH.

ofdkit invoice 的输出示例（数据是虚构的）/ Example output (fictional data):

{
  "invoice_number": "00000000000000000001",
  "invoice_code": null,
  "issue_date": "2026-01-02",
  "check_code": null,
  "machine_number": null,
  "buyer": {"name": "测试单位甲", "tax_id": "TESTBUYER0001", "address_phone": null, "bank_account": null},
  "seller": {"name": "测试单位乙", "tax_id": "TESTSELLER0002", "address_phone": null, "bank_account": null},
  "amount_without_tax": "100.00",
  "tax_amount": "6.00",
  "total_amount": "106.00",
  "issuer": "测试人员",
  "payee": null,
  "reviewer": null,
  "note": null,
  "items": [
    {"name": "*测试*服务项目", "specification": null, "unit": null, "quantity": null,
     "unit_price": null, "amount": "100.00", "tax_rate": "6%", "tax_amount": "6.00"}
  ],
  "sources": ["custom_tags"],
  "warnings": []
}

金额是字符串，不会丢失精度。如果"合计金额 + 合计税额"不等于"价税合计"，warnings 里会提示。 Amounts are decimal strings, so no precision is lost. warnings notes totals that do not add up.

在 AI 工具里使用 / Use from AI tools

MCP 服务器 / MCP server

让 Claude、Codex、Cursor 等 AI 工具直接读取 OFD 文件。需要先安装 uv。所有工具都是只读的。 Lets Claude, Codex, Cursor and other AI tools read OFD files. Requires uv. All tools are read-only.

Claude Code:

claude mcp add ofdkit -- uvx --from "ofdkit[mcp]" ofdkit mcp

Claude Desktop、Cursor 等使用 JSON 配置的工具 / tools with a JSON config:

{
  "mcpServers": {
    "ofdkit": {
      "command": "uvx",
      "args": ["--from", "ofdkit[mcp]", "ofdkit", "mcp"]
    }
  }
}

Codex（~/.codex/config.toml）:

[mcp_servers.ofdkit]
command = "uvx"
args = ["--from", "ofdkit[mcp]", "ofdkit", "mcp"]

加上 --root 文件夹 可以只允许读取这个文件夹里的文件。 Add --root FOLDER to allow only files inside that folder.

Agent Skill

skills/ofd-reader 告诉 AI 什么时候、怎么用 ofdkit 命令行，适用于支持 Agent Skills 的工具。把这个文件夹复制到： skills/ofd-reader tells an AI agent when and how to use the ofdkit command line. Copy the folder to:

• Claude Code: ~/.claude/skills/ofd-reader/
• Codex: ~/.codex/skills/ofd-reader/

MarkItDown 插件 / MarkItDown plugin

MarkItDown 是把各种文件转成 Markdown 给 AI 读的工具。安装插件后它就能读 OFD： MarkItDown converts files to Markdown for AI. With this plugin it reads OFD too:

pip install "ofdkit[markitdown]"
markitdown --use-plugins 发票.ofd

Python

import ofdkit

with ofdkit.open("发票.ofd") as doc:
    print(doc.text())
    invoice = doc.invoice()        # 不是发票时返回 None / None if not an invoice
    if invoice:
        print(invoice.invoice_number, invoice.total_amount)

# 文件夹、.zip 压缩包或多个文件 / folders, ZIP archives or several files
for result in ofdkit.read_invoices("发票文件夹", recursive=True, buyer=["某某有限公司"]):
    if result.invoice:
        print(result.file, result.invoice.total_amount, "重复" if result.duplicate else "", result.buyer_mismatch)
    else:
        print(result.file, result.error)

遇到读不出的文件 / Reporting a problem

请不要把真实发票发到网上。用这个命令导出文件结构，文字会全部打码，只保留结构和位置，然后附在 GitHub issue 里： Please do not post real invoices. Export the structure with all text masked and attach it to a GitHub issue:

ofdkit describe 文件.ofd -o 结构.json

注意 / Notes

• ofdkit 只读取文件里写了什么，不验证发票真伪。查验真伪请使用国家税务总局的发票查验平台。 ofdkit reports what a file contains. It does not verify that an invoice is genuine.
• 文字如果是画成图形或图片的，就读不出来（需要 OCR，不在本项目范围内）。 Text drawn as shapes or images cannot be extracted (that needs OCR, which is out of scope).
• 阅读顺序是根据位置推测的。两栏的正文会按栏读，表格按行读；版面特别复杂的页面（比如多栏加图片的杂志页）可能不完全准确。 Reading order is inferred from positions. Prose columns are read column by column and tables row by row; very complex pages (magazine layouts with pictures) may not come out perfectly.

原则 / Principles

• 只读取 OFD，不生成、不修改发票。 Reads OFD only. It never creates or modifies invoices.
• 本项目大部分代码由 AI（Claude）编写，维护者负责决定方向和批准发布。 Most code is written with AI assistance (Claude). The maintainer sets the direction and approves releases.

许可证 / License

Apache-2.0