AI-powered terminal assistant for LaTeX academic papers

Installation · Quick Start · Commands · Configuration · Guide

TexGuardian is a researcher-focused interactive CLI that helps you write, verify, and polish LaTeX academic papers for conference submission. Think of it as Claude Code, but for LaTeX — it reads your entire .tex and .bib files, understands venue requirements, and fixes issues through reviewable unified diff patches with checkpoint safety.

It runs a 7-step review pipeline in a single command: compile your paper, run verification checks, fix issues with LLM-generated patches, validate citations against CrossRef and Semantic Scholar, analyze figures and tables, and run a visual polish loop that renders your PDF and sends pages to a vision model to catch layout problems text analysis can't see. Every edit is shown as a diff you approve. Automatic checkpoints mean you can always roll back.

Whether you're preparing a NeurIPS submission, anonymizing for double-blind review, converting to camera-ready, or validating that every citation in your bibliography actually exists — TexGuardian handles the tedious parts so you can focus on the research.

• Styled REPL — bordered welcome panel with paper stats, colored ❯ prompt
• 26 slash commands for every stage of paper preparation
• LLM-powered fixes — figures, tables, citations, sections, anonymization, camera-ready
• Instant verification — regex-based checks run directly on .tex files, no compilation needed
• Unified diff patches — every edit is reviewable, with checkpoints and rollback
• Visual polish loop — renders PDF, sends pages to vision model, fixes layout issues
• Natural language — ask for anything in plain English, the LLM figures out the rest
• Two providers — AWS Bedrock or OpenRouter, switch with one command

pip install texguardian

git clone https://github.com/arcAman07/TexGuardian.git
cd TexGuardian
pip install -e ".[dev]"

TexGuardian needs LaTeX and Poppler installed on your system for compilation and visual checks:

Option A: TinyTeX (~250 MB) — recommended:

# macOS / Linux
curl -sL "https://yihui.org/tinytex/install-bin-unix.sh" | sh

# Add to PATH (add to your ~/.zshrc or ~/.bashrc):
export PATH="$HOME/Library/TinyTeX/bin/universal-darwin:$PATH"  # macOS
export PATH="$HOME/.TinyTeX/bin/x86_64-linux:$PATH"             # Linux

# Install common academic packages:
tlmgr install booktabs natbib hyperref pgfplots xcolor float geometry \
              amsmath amssymb graphicx tikz caption subcaption

TinyTeX is a minimal, portable TeX Live distribution. Missing packages are auto-installed on first compile, or install manually with tlmgr install <pkg>. See yihui.org/tinytex for details.

Option B: Full TeX Live (~4 GB) — includes every package, no missing-package issues:

# macOS
brew install --cask mactex-no-gui

# Ubuntu/Debian
sudo apt install texlive-full

texguardian doctor

TexGuardian Doctor

Checking external tools...

  ✓ latexmk      ~/Library/TinyTeX/bin/universal-darwin/latexmk
  ✓ pdflatex     ~/Library/TinyTeX/bin/universal-darwin/pdflatex
  ✓ pdfinfo      /opt/homebrew/bin/pdfinfo
  ✓ pdftoppm     /opt/homebrew/bin/pdftoppm

All tools found. You're good to go!

cd /path/to/your/latex/paper
texguardian init

This creates three files:

AWS Bedrock (recommended):

# Create a .env file in your project root:
AWS_ACCESS_KEY_ID=your-key
AWS_SECRET_ACCESS_KEY=your-secret
AWS_REGION=us-east-1

OpenRouter:

# In texguardian.yaml:
providers:
  default: "openrouter"
  openrouter:
    api_key: "sk-or-..."  # Get from https://openrouter.ai/keys

texguardian chat

On startup you'll see a styled welcome panel with paper stats:

╭──────────────────────────────────────────────────────────╮
│  TexGuardian                                             │
│                                                          │
│  Paper  Scaling Sparse MoE for Long-Context Doc...       │
│  Venue  NeurIPS 2026          Deadline  2026-05-15       │
│  Model  claude opus 4.5       Provider  bedrock          │
│  File   demo_paper.tex        Figures 4 · Tables 3       │
│                                                          │
│  Type /help for commands or ask a question.              │
╰──────────────────────────────────────────────────────────╯

❯ _

Then interact with commands or plain English:

❯ /figures fix
Figure Pipeline

Step 1: Verifying Figures
  Found 4 figures, 3 issues detected

Step 2: LLM Fix
  Analyzing figure issues...

  ✓ Generated 3 patches:
    1. demo_paper.tex (+1/-1) — width=1.3\columnwidth → width=\columnwidth
    2. demo_paper.tex (+1/-1) — width=1.8\textwidth → width=\columnwidth
    3. demo_paper.tex (+1/-1) — width=2.0\columnwidth → width=\columnwidth

  [A]pply all | [R]eview | [N]o: a
  ✓ Checkpoint created
  ✓ All patches applied

❯ The tables use \hline — please fix them to use booktabs

You can also just type in plain English. TexGuardian will understand:

❯ The figure on line 44 has width=1.3\columnwidth which causes overflow. Please fix it.
❯ Please make this paper anonymous for double-blind review
❯ Can you suggest more citations for the related work section?

The /venue and /model commands also accept natural language:

❯ /venue please download the neurips 2026 style files
❯ /model we need to switch to claude sonnet 4 on bedrock

project:
  main_tex: "main.tex"          # Your main .tex file
  output_dir: "build"           # Output directory for PDF

providers:
  default: "bedrock"            # "bedrock" or "openrouter"
  bedrock:
    region: "us-east-1"
    # Credentials loaded from .env automatically
  openrouter:
    api_key: "sk-or-..."
    base_url: "https://openrouter.ai/api/v1"

models:
  default: "claude opus 4.5"    # Main model for analysis and fixes
  vision: "claude opus 4.5"     # Vision model for visual polish

safety:
  max_changed_lines: 50         # Max lines per patch
  max_rounds: 10                # Max auto-fix iterations
  max_visual_rounds: 5          # Max visual verification loops
  allowlist:                    # Files the LLM can modify
    - "*.tex"
    - "*.bib"
    - "*.sty"
    - "*.cls"
  denylist:                     # Files the LLM cannot touch
    - ".git/**"
    - "*.pdf"
    - "build/**"

latex:
  compiler: "latexmk"           # Build tool
  engine: "pdflatex"            # TeX engine
  shell_escape: false           # Enable --shell-escape
  timeout: 240                  # Compilation timeout (seconds)

visual:
  dpi: 150                      # PDF render resolution
  diff_threshold: 5.0           # Convergence threshold (%)
  pixel_threshold: 15           # Per-pixel diff threshold (0-255)
  max_pages_to_analyze: 0       # 0 = all pages

The paper spec defines venue-specific rules, custom checks, and your system prompt:

---
title: "Your Paper Title"
venue: "NeurIPS 2026"
deadline: "2026-05-15"
thresholds:
  max_pages: 9
  min_references: 30
  max_self_citation_ratio: 0.2
human_review:
  - "Changes to abstract"
  - "Deletion of more than 10 lines"
---

## Custom Checks

```check
name: citation_format
severity: warning
pattern: "\\cite\\{(?!p|t)"
message: "Use \\citep{} or \\citet{} instead of \\cite{}"
```

```check
name: todo_remaining
severity: error
pattern: "TODO|FIXME|XXX"
message: "Remove TODO/FIXME markers before submission"
```

## System Prompt

```system-prompt
You are an expert academic writing assistant specializing in machine learning.
Use formal academic English. Every claim must be backed by evidence.
```

Plus any model from openrouter.ai/models.

# Run tests
pytest tests/ -v

# Run tests with coverage
pytest tests/ --cov=texguardian --cov-report=term-missing

# Lint
ruff check src/

# Type check
mypy src/texguardian

See docs/GUIDE.md for the full architecture and project layout.

MIT License. See LICENSE for details.