anytomd

A pure Rust tool and library that converts various document formats into Markdown — designed for LLM consumption.

Why?

MarkItDown is a great Python library for converting documents to Markdown. But integrating Python into Rust applications means bundling a Python runtime (~50 MB), dealing with cross-platform compatibility issues, and managing dependency hell.

anytomd solves this with a single cargo add anytomd — zero external runtime, no C bindings, no subprocess calls. Just pure Rust.

Supported Formats

| Format | Extensions | Notes | |--------|-----------|-------| | DOCX | .docx | Headings, tables, lists, bold/italic, hyperlinks, images, text boxes | | PPTX | .pptx | Slides, tables, speaker notes, images, group shapes | | XLSX | .xlsx | Multi-sheet, date/time handling, images | | XLS | .xls | Legacy Excel (via calamine) | | HTML | .html, .htm | Full DOM: headings, tables, lists, links, blockquotes, code blocks | | CSV | .csv | Converted to Markdown tables | | Jupyter Notebook | .ipynb | Markdown cells preserved, code cells in fenced blocks with language detection | | JSON | .json | Pretty-printed in fenced code blocks | | XML | .xml | Pretty-printed in fenced code blocks | | Images | .png, .jpg, .gif, .webp, .bmp, .tiff, .svg, .heic, .avif | Optional LLM-based alt text via ImageDescriber | | Code | .py, .rs, .js, .ts, .c, .cpp, .go, .java, .rb, .swift, .sh, ... | Fenced code blocks with language identifier | | Plain Text | .txt, .md, .rst, .log, .toml, .yaml, .ini, etc. | Passthrough with encoding detection (UTF-8, UTF-16, Windows-1252) |

Note on PDF: PDF conversion is intentionally out of scope. Gemini, ChatGPT, and Claude already provide native PDF support (with plan/model-specific limits), so anytomd focuses on formats that still benefit from dedicated Markdown conversion. Attempting to convert a PDF will return a descriptive FormatNotSupported error.

Format is auto-detected from magic bytes and file extension. ZIP-based formats (DOCX/PPTX/XLSX) are distinguished by inspecting internal archive structure.

Conversion Examples

CSV

A CSV file with multilingual data:

Name,Age,City
Alice,30,Seoul
Bob,25,東京
Charlie,35,New York
다영,28,서울

Output:

| Name | Age | City |
|---|---|---|
| Alice | 30 | Seoul |
| Bob | 25 | 東京 |
| Charlie | 35 | New York |
| 다영 | 28 | 서울 |

DOCX

A Word document with headings, links, Korean text, and emoji:

Output:

# Sample Document

This is a simple paragraph.

## Section One

Visit [Example](https://example.com) for more info.

Korean: 한국어 테스트

Emoji: 🚀✨🌍

### Subsection

Final paragraph with mixed content.

PPTX

A PowerPoint presentation with slides, tables, speaker notes, and multilingual content:

Output:

## Slide 1: Sample Presentation

Welcome to the presentation.

---

## Slide 2

Data Overview

| Name | Value | Status |
|---|---|---|
| Alpha | 100 | Active |
| Beta | 200 | Inactive |
| Gamma | 300 | Active |

> Note: Remember to explain the data table.

---

## Slide 3: Multilingual

한국어 테스트
🚀✨🌍

> Note: Test multilingual rendering.

Installation

Rust (Cargo)

cargo add anytomd

npm (WASM)

npm install anytomd

import init, { convertBytes } from 'anytomd';

await init();

const response = await fetch('document.docx');
const bytes = new Uint8Array(await response.arrayBuffer());

const result = convertBytes(bytes, 'docx');
console.log(result.markdown);

Feature Flags

| Feature | Dependencies | Description | |---------|-------------|-------------| | (default) | async-gemini | Async API + AsyncGeminiDescriber — all async features enabled out of the box | | async | futures-util | Async API (convert_file_async, convert_bytes_async, AsyncImageDescriber trait) | | async-gemini | async + reqwest | AsyncGeminiDescriber for concurrent image descriptions via Gemini | | wasm | wasm-bindgen, js-sys, wasm-bindgen-futures | WebAssembly bindings (convertBytes, convertBytesWithOptions) for browser/edge use | | wasm + async-gemini | (combined) | Adds convertBytesWithGemini for async Gemini-powered conversion in WASM |

Async features are included by default. To opt out:

anytomd = { version = "1", default-features = false }

WebAssembly (WASM)

anytomd compiles to wasm32-unknown-unknown, enabling client-side document conversion in browsers, Cloudflare Workers, Deno Deploy, and other edge runtimes. Documents never leave the user's device.

Build

# Basic WASM build (sync conversion only)
wasm-pack build --target web --no-default-features --features wasm

# With Gemini async image descriptions
wasm-pack build --target web --no-default-features --features wasm,async-gemini

Usage from JavaScript

import init, { convertBytes } from './pkg/anytomd.js';

await init();

const response = await fetch('document.docx');
const bytes = new Uint8Array(await response.arrayBuffer());

const result = convertBytes(bytes, 'docx');
console.log(result.markdown);
console.log(result.plainText);
console.log(result.title);       // string or null
console.log(result.warnings);    // string[]

With Gemini Image Descriptions (requires wasm + async-gemini features)

import init, { convertBytesWithGemini } from './pkg/anytomd.js';

await init();

const response = await fetch('presentation.pptx');
const bytes = new Uint8Array(await response.arrayBuffer());

// Images are described concurrently via the Gemini API
const result = await convertBytesWithGemini(bytes, 'pptx', 'your-gemini-api-key');
console.log(result.markdown);  // images have LLM-generated alt text

WASM API Availability

| API | Native | WASM | |-----|--------|------| | convert_bytes / convertBytes | Yes | Yes | | convert_bytes_async | Yes | Yes | | convert_file / convert_file_async | Yes | No (no filesystem) | | GeminiDescriber (sync) | Yes | No (uses ureq) | | AsyncGeminiDescriber / convertBytesWithGemini | Yes | Yes (wasm + async-gemini) |

All 12 format converters work on WASM via convert_bytes.

CLI

Install

cargo install anytomd

Usage

# Convert a single file
anytomd document.docx > output.md

# Convert multiple files (separated by <!-- source: path --> comments)
anytomd report.docx data.csv slides.pptx > combined.md

# Write output to a file
anytomd document.docx -o output.md

# Read from stdin (--format is required)
cat data.csv | anytomd --format csv

# Override format detection
anytomd --format html page.dat

# Strict mode: treat recoverable errors as hard errors
anytomd --strict document.docx

# Plain text output (Markdown formatting stripped)
anytomd --plain-text document.docx

# Plain text from stdin
echo "Name,Age" | anytomd --format csv --plain-text

# Auto image descriptions (just set GEMINI_API_KEY)
export GEMINI_API_KEY=your-key
anytomd presentation.pptx

Exit Codes

| Code | Meaning | |------|---------| | 0 | Success | | 1 | Conversion failure | | 2 | Invalid arguments |

Quick Start (Library)

use anytomd::{convert_file, convert_bytes, ConversionOptions};

// Convert a file (format auto-detected from extension and magic bytes)
let options = ConversionOptions::default();
let result = convert_file("document.docx", &options).unwrap();
println!("{}", result.markdown);

// Convert raw bytes with an explicit format
let csv_data = b"Name,Age\nAlice,30\nBob,25";
let result = convert_bytes(csv_data, "csv", &options).unwrap();
println!("{}", result.markdown);

Plain Text Output

Every conversion produces both Markdown and plain text output. The plain text is extracted directly from the source document — no post-processing or markdown stripping — so source characters like **kwargs or # comment are preserved exactly.

use anytomd::{convert_file, ConversionOptions};

let result = convert_file("document.docx", &ConversionOptions::default()).unwrap();

// Markdown output
println!("{}", result.markdown);

// Plain text output (no headings, bold, tables, code fences, etc.)
println!("{}", result.plain_text);

Extracting Embedded Images

use anytomd::{convert_file, ConversionOptions};

let options = ConversionOptions {
    extract_images: true,
    ..Default::default()
};
let result = convert_file("presentation.pptx", &options).unwrap();

for (filename, bytes) in &result.images {
    std::fs::write(filename, bytes).unwrap();
}

LLM-Based Image Descriptions

anytomd can generate alt text for images using any LLM backend via the ImageDescriber trait. A built-in Google Gemini implementation is included.

use std::sync::Arc;
use anytomd::{convert_file, ConversionOptions, ImageDescriber, ConvertError};
use anytomd::gemini::GeminiDescriber;

// Option 1: Use the built-in Gemini describer
let describer = GeminiDescriber::from_env()  // reads GEMINI_API_KEY
    .unwrap()
    .with_model("gemini-3-flash-preview".to_string());

let options = ConversionOptions {
    image_describer: Some(Arc::new(describer)),
    ..Default::default()
};
let result = convert_file("document.docx", &options).unwrap();
// Images now have LLM-generated alt text: ![A chart showing quarterly revenue](chart.png)

// Option 2: Implement your own describer for any backend
struct MyDescriber;

impl ImageDescriber for MyDescriber {
    fn describe(
        &self,
        image_bytes: &[u8],
        mime_type: &str,
        prompt: &str,
    ) -> Result<String, ConvertError> {
        // Call your preferred LLM API here
        Ok("description of the image".to_string())
    }
}

Async Image Descriptions

For documents with many images, the async API resolves all descriptions concurrently. Included by default since v0.11.0.

use std::sync::Arc;
use anytomd::{convert_file_async, AsyncConversionOptions, AsyncImageDescriber, ConvertError};
use anytomd::gemini::AsyncGeminiDescriber;

#[tokio::main]
async fn main() {
    let describer = AsyncGeminiDescriber::from_env().unwrap();

    let options = AsyncConversionOptions {
        async_image_describer: Some(Arc::new(describer)),
        ..Default::default()
    };

    let result = convert_file_async("presentation.pptx", &options).await.unwrap();
    println!("{}", result.markdown);
    // All images described concurrently — significant speedup for multi-image documents
}

The library has no tokio dependency — the caller provides the async runtime. Any runtime (tokio, async-std, etc.) works.

API

convert_file

/// Convert a file at the given path to Markdown.
/// Format is auto-detected from magic bytes and file extension.
pub fn convert_file(
    path: impl AsRef<Path>,
    options: &ConversionOptions,
) -> Result<ConversionResult, ConvertError>

convert_bytes

/// Convert raw bytes to Markdown with an explicit format extension.
pub fn convert_bytes(
    data: &[u8],
    extension: &str,
    options: &ConversionOptions,
) -> Result<ConversionResult, ConvertError>

convert_file_async

Included by default (requires the async feature if default features are disabled).

/// Convert a file at the given path to Markdown with async image description.
/// If an async_image_describer is set, all image descriptions are resolved concurrently.
pub async fn convert_file_async(
    path: impl AsRef<Path>,
    options: &AsyncConversionOptions,
) -> Result<ConversionResult, ConvertError>

convert_bytes_async

Included by default (requires the async feature if default features are disabled).

/// Convert raw bytes to Markdown with async image description.
pub async fn convert_bytes_async(
    data: &[u8],
    extension: &str,
    options: &AsyncConversionOptions,
) -> Result<ConversionResult, ConvertError>

ConversionOptions

| Field | Type | Default | Description | |-------|------|---------|-------------| | extract_images | bool | false | Extract embedded images into result.images | | max_total_image_bytes | usize | 50 MB | Hard cap for total extracted image bytes | | max_input_bytes | usize | 100 MB | Maximum input file size | | max_uncompressed_zip_bytes | usize | 500 MB | ZIP bomb guard | | strict | bool | false | Error on recoverable failures instead of warnings | | image_describer | Option<Arc<dyn ImageDescriber>> | None | LLM backend for image alt text generation |

ConversionResult

pub struct ConversionResult {
    pub markdown: String,                  // The converted Markdown
    pub plain_text: String,                // Plain text (extracted directly, no markdown syntax)
    pub title: Option<String>,             // Document title, if detected
    pub images: Vec<(String, Vec<u8>)>,    // Extracted images (filename, bytes)
    pub warnings: Vec<ConversionWarning>,  // Recoverable issues encountered
}

Error Handling

Conversion is best-effort by default. If a single element fails to parse (e.g., a corrupted table), it is skipped and a warning is added to result.warnings. The rest of the document is still converted.

Set strict: true in ConversionOptions to turn recoverable failures into errors instead.

Warning codes: SkippedElement, UnsupportedFeature, ResourceLimitReached, MalformedSegment.

Development

Build and Test

cargo build && cargo test && cargo clippy -- -D warnings

Docker

A Docker environment is available for reproducible Linux builds:

docker compose run --rm verify    # Full loop: fmt + clippy + test + release build
docker compose run --rm test      # Run all tests
docker compose run --rm lint      # clippy + fmt check
docker compose run --rm shell     # Interactive bash

License

Apache-2.0