Nougat-MCP

nougat-mcp is a Model Context Protocol (MCP) server for high-fidelity OCR of scientific PDFs using Meta's Nougat.

It is designed for agent workflows where you need equations, tables, and structure preserved better than traditional OCR.

Why This Server

• Scientific OCR quality tailored for papers, formulas, and dense layouts.
• MCP-native interface for Codex, Claude, Cursor, Antigravity, and other clients.
• Output-format control:
  • mmd: raw Nougat/Mathpix-style output.
  • md: renderer-friendly conversion (math delimiter and KaTeX compatibility fixes).
• Settings file support so agents can read a shared default format policy.

Installation

Install from PyPI:

uv pip install nougat-mcp

This package installs nougat-ocr and pins known-sensitive dependencies for stability.

Tools

parse_research_paper

Arguments:

• file_path (string): Absolute path to a local PDF.
• output_format (string, optional):
  • default (default): uses server settings.
  • mmd: raw Nougat output.
  • md: converted markdown-friendly output.

Returns:

• OCR result as a single text string in the requested format.

get_output_settings

Returns resolved server output settings, including where settings were loaded from.

Output Conversion (mmd -> md)

When output_format="md", the server applies compatibility conversions:

• \[ ... \] -> $$ ... $$
• \( ... \) -> $ ... $
• \tag{...} -> visible equation label \qquad\text{(...)}
• KaTeX delimiter normalization, for example:
  • \bigl{\|} ... \bigr{\|} -> \bigl\| ... \bigr\|

This avoids common renderer parse errors in markdown environments that are not fully MathJax-compatible.

Server Settings

Settings are read in this order:

1. NOUGAT_MCP_SETTINGS (if set)
2. ./settings.json (current working directory)

Example settings.json:

{
  "nougat_mcp": {
    "default_output_format": "md",
    "md_rewrite_tags": true,
    "md_fix_sized_delimiters": true
  }
}

Agent Configuration

Codex CLI

Add to ~/.codex/config.toml:

[mcp_servers.nougat]
command = "uvx"
args = ["nougat-mcp"]
enabled = true

[mcp_servers.nougat.env]
NOUGAT_MCP_SETTINGS = "/absolute/path/to/settings.json"

Claude Desktop

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "nougat": {
      "command": "uvx",
      "args": ["nougat-mcp"],
      "env": {
        "NOUGAT_MCP_SETTINGS": "/absolute/path/to/settings.json"
      }
    }
  }
}

Antigravity / Gemini Desktop

Add to ~/.gemini/settings.json:

{
  "mcpServers": {
    "nougat": {
      "type": "stdio",
      "command": "uvx",
      "args": ["nougat-mcp"],
      "env": {
        "NOUGAT_MCP_SETTINGS": "/absolute/path/to/settings.json"
      }
    }
  }
}

Cursor

In Cursor MCP settings, add:

{
  "mcpServers": {
    "nougat": {
      "command": "uvx",
      "args": ["nougat-mcp"],
      "env": {
        "NOUGAT_MCP_SETTINGS": "/absolute/path/to/settings.json"
      }
    }
  }
}

Note: Cursor MCP config location can vary by version/platform; use the MCP settings UI or your current JSON settings file.

Showcase (Real Page Example)

A real extraction from page 5 of src/2405.08770v1.pdf is included:

• Input PDF page: showcase/2405.08770v1_page5.pdf
• Raw mmd output: showcase/2405.08770v1_page5.mmd
• Converted md output: showcase/2405.08770v1_page5.md

Quick comparison:

# mmd
\[DV=V_{x}. \tag{3.2}\]

# md
$$
DV=V_{x}. \qquad\text{(3.2)}
$$

Performance Notes

• First run may download model weights (~1.4 GB).
• CPU inference is significantly slower than GPU inference.
• Use page subsets whenever possible to reduce runtime.

Compatibility Pins

To keep Nougat stable across environments, the package pins sensitive dependency ranges:

• transformers>=4.35,<4.38
• albumentations>=1.3,<1.4
• pypdfium2<5.0
• huggingface-hub<1.0
• fsspec<=2025.10.0

Credits

• Nougat OCR: https://github.com/facebookresearch/nougat
• Paper: https://arxiv.org/abs/2308.13418

License

GNU General Public License v3.0 (LICENSE).