一个带有MCP接口的LaTeX文档编译器。AI代理运行在结构化文档模型上——包括章节、段落、图表——而TeXFlow则处理所有LaTeX机制：包、前言、字体和编译。

安装

或者不安装：

系统依赖（可选）

TeXFlow 使用 XeLaTeX 将文档编译成 PDF。没有它，你仍然可以构建和导出 .tex 文件。

页面预览（PNG）时，安装 poppler-utils（提供 pdftoppm。

用Claude代码配置

就这些。重启Claude代码，工具就能用了。

要设置工作区目录（存放文档的地方）：

使用 Claude Desktop 配置

补充你的claude_desktop_config.json：

它是如何运作的

TeXFlow 有两个条目流，都汇聚在同一内存文档模型上：

1. 脚手架 — document(action="create") 构建一个空白的文档骨架。用 edit(action="insert") 添加内容。
2. Markdown 导入 — document(action="ingest", source="paper.md") 将 markdown 解析到模型中。从这里细化布局。

模型会自动保存为 JSON 格式到磁盘。LaTeX 永远只是输出伪造物——你永远不会直接编辑 .tex。

工具

TeXFlow MCP

A LaTeX document compiler with an MCP interface. AI agents operate on a structured document model — sections, paragraphs, figures, tables — while TeXFlow handles all LaTeX mechanics: packages, preamble, fonts, and compilation.

Install

pip install texflow-mcp

Or run without installing:

uvx texflow-mcp

System dependencies (optional)

TeXFlow compiles documents to PDF using XeLaTeX. Without it, you can still build and export .tex files.

# Arch
pacman -S texlive-xetex texlive-fontsrecommended texlive-fontsextra

# Debian/Ubuntu
apt install texlive-xetex texlive-fonts-recommended texlive-fonts-extra

# Fedora
dnf install texlive-xetex texlive-collection-fontsrecommended texlive-collection-fontsextra

For page preview (PNG), install poppler-utils (provides pdftoppm).

Configure with Claude Code

claude mcp add texflow -- uvx texflow-mcp

That’s it. Restart Claude Code and the tools are available.

To set a workspace directory (where documents are saved):

claude mcp add texflow -- uvx texflow-mcp ~/Documents/TeXFlow

Configure with Claude Desktop

Add to your claude_desktop_config.json:

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

How it works

TeXFlow has two entry flows, both converging on the same in-memory document model:

1. Scaffold — document(action="create") builds an empty document skeleton. Add content with edit(action="insert").
2. Markdown ingest — document(action="ingest", source="paper.md") parses markdown into the model. Refine layout from there.

The model auto-saves to disk as JSON. LaTeX is only ever an output artifact — you never edit .tex directly.

Tools

Every response includes a workflow state hint showing where you are and what to do next.

Example session

> document(action="create", title="My Paper", document_class="article")

> queue(operations=[
    {"tool": "edit", "action": "insert", "block_type": "section", "title": "Introduction", "level": 1},
    {"tool": "edit", "action": "insert", "content": "This paper explores...", "section": "Introduction"},