---
title: Markdown Layout Extractor
emoji: 📄
colorFrom: red
colorTo: yellow
sdk: docker
app_port: 7860
pinned: false
---

<p>
  <div align="center">
  <h1>
    PDF to Markdown MCP
    <br /> <br />
    <a href="">
      <img
        src="https://img.shields.io/badge/python%20%7C%203.12-blue"
        alt="Python 3.12"
      />
    </a>
    <a href="https://github.com/astral-sh/uv">
      <img
        src="https://img.shields.io/endpoint?url=https://raw.githubusercontent.com/astral-sh/uv/main/assets/badge/v0.json"
        alt="uv"
      />
    </a>
    <a href="https://modelcontextprotocol.io/">
      <img
        src="https://img.shields.io/badge/MCP-FastMCP-6C47FF"
        alt="FastMCP"
      />
    </a>
    <a href="https://mistral.ai/">
      <img
        src="https://img.shields.io/badge/Mistral%20AI-FF7000?logoColor=white"
        alt="Mistral AI"
      />
    </a>
    <a href="https://www.starlette.io/">
      <img
        src="https://img.shields.io/badge/Starlette-ASGI-009688"
        alt="Starlette"
      />
    </a>
    <a href="https://www.uvicorn.org/">
      <img
        src="https://img.shields.io/badge/Uvicorn-server-4051B5"
        alt="Uvicorn"
      />
    </a>
    <a href="https://loguru.readthedocs.io/">
      <img
        src="https://img.shields.io/badge/Loguru-logging-FF6B6B"
        alt="Loguru"
      />
    </a>
  </h1>
  </div>
</p>

An MCP (Model Context Protocol) server that converts PDFs and documents into Markdown using **Mistral OCR**.

## Features

- **`pdf_to_markdown`** — Convert any publicly accessible PDF/document URL to merged Markdown.
- **`pdf_to_structured_markdown`** — Convert and get per-page structured output (page index, individual markdown, merged result).
- CORS-enabled SSE transport — connect from any MCP client or inspector.
- `/health` endpoint for liveness probing.
- Structured, colorized logging via Loguru.

## Project Structure

```
pdf_to_md_mcp/
├── main.py                       # Entry point — uvicorn runner
├── pyproject.toml
├── sample.env                    # Secrets reference (copy to .env)
├── development.yml               # Non-secret config (server, CORS, OCR model)
└── app/
    ├── server.py                 # ASGI app factory (MCP + CORS + health)
    ├── core/
    │   ├── config.py             # Pydantic settings (loads .env + development.yml)
    │   ├── logger.py             # Loguru logger
    │   ├── lifespan.py           # AppContext + Mistral client lifecycle
    │   └── exceptions.py         # Domain exceptions
    ├── services/
    │   └── ocr_service.py        # Mistral OCR business logic
    ├── tools/
    │   └── markdown_tools.py     # @mcp.tool() definitions
    └── utils/
        ├── response.py           # create_response() helper
        └── validators.py         # URL validation
```

## Setup

```bash
# Install uv if not already installed
curl -LsSf https://astral.sh/uv/install.sh | sh

# Install dependencies
uv sync

# Configure secrets
cp sample.env .env
# Edit .env — set MISTRAL_API_KEY
# Non-secret config (server, CORS, OCR model) lives in development.yml
```

## Run

```bash
uv run main.py
```

Server starts at `http://127.0.0.1:8000` by default.

| Endpoint | Description |
| --- | --- |
| `GET /health` | Liveness probe |
| `GET /sse` | MCP SSE transport |
| `POST /messages/` | MCP message handler |

## MCP Tools

### `pdf_to_markdown`

Convert a document URL to merged Markdown (all pages concatenated).

**Input**

| Parameter | Type | Description |
| --- | --- | --- |
| `document_url` | `string` | Publicly accessible URL of a PDF or image document |

**Returns** — `string`

```
# Introduction

This paper presents...

## Section 2

...
```

---

### `pdf_to_structured_markdown`

Convert a document URL and get per-page structured output alongside the merged result.

**Input**

| Parameter | Type | Description |
| --- | --- | --- |
| `document_url` | `string` | Publicly accessible URL of a PDF or image document |

**Returns** — `object`

```json
{
  "page_count": 3,
  "pages": [
    { "index": 0, "markdown": "# Page 1\n..." },
    { "index": 1, "markdown": "## Page 2\n..." },
    { "index": 2, "markdown": "### Page 3\n..." }
  ],
  "markdown": "# Page 1\n...\n\n## Page 2\n...\n\n### Page 3\n..."
}
```

## Debugging with MCP Inspector

```bash
npx -y @modelcontextprotocol/inspector
```

Connect to `http://127.0.0.1:8000/sse` locally or your Railway URL in production.

## Deploy to Railway

### 1. Push to GitHub

```bash
git init
git add .
git commit -m "initial commit"
gh repo create pdf-to-md-mcp --public --source=. --push
```

### 2. Create a Railway project

Go to [railway.app](https://railway.app) → **New Project** → **Deploy from GitHub repo** → select your repo.

Railway detects the `railway.json` and uses `uv run main.py` as the start command automatically.

### 3. Set environment variables

In Railway → your service → **Variables**, add:

| Variable | Value |
|---|---|
| `MISTRAL_API_KEY` | your Mistral API key |
| `HOST` | `0.0.0.0` |

> `PORT` is injected automatically by Railway — do **not** set it manually.  
> All other config (`MISTRAL_OCR_MODEL`, `LOG_LEVEL`, etc.) is read from `development.yml`.

### 4. Deploy

Railway triggers a deploy on every push to your default branch. Once live, your public SSE URL will be:

```
https://<your-service>.up.railway.app/sse
```

Use that URL in any MCP client or pass it to the inspector:

```bash
npx -y @modelcontextprotocol/inspector
# connect to: https://<your-service>.up.railway.app/sse
```

### Why it works

- Railway injects `PORT` as an env var — pydantic-settings reads env vars before `development.yml`, so it's picked up automatically.
- `HOST=0.0.0.0` (set via Railway Variables) overrides the local `127.0.0.1` default so the container is reachable.
- `proxy_headers=True` in `main.py` makes uvicorn trust Railway's `X-Forwarded-*` headers.
- `/health` is set as Railway's healthcheck path in `railway.json`.



## Configuration

Configuration is split across two files to separate secrets from non-sensitive settings.

### `.env` — Secrets only

```dotenv
MISTRAL_API_KEY=your_mistral_api_key_here
```

### `development.yml` — Non-secret config

```yaml
# Mistral
MISTRAL_OCR_MODEL: mistral-ocr-latest
MISTRAL_TABLE_FORMAT: markdown

# Server
APP_NAME: "Markdown & Layout Extractor"
HOST: "127.0.0.1"
PORT: 8000
LOG_LEVEL: INFO

# CORS
CORS_ALLOW_ORIGINS:
  - "*"
CORS_ALLOW_METHODS:
  - "*"
CORS_ALLOW_HEADERS:
  - "*"
```

**Priority (highest → lowest):** environment variables → `.env` → `development.yml`

### All settings

| Variable | File | Default | Description |
| --- | --- | --- | --- |
| `MISTRAL_API_KEY` | `.env` | **required** | Mistral AI API key |
| `MISTRAL_OCR_MODEL` | `development.yml` | `mistral-ocr-latest` | OCR model identifier |
| `MISTRAL_TABLE_FORMAT` | `development.yml` | `markdown` | Table output format |
| `APP_NAME` | `development.yml` | `Markdown & Layout Extractor` | MCP server name |
| `HOST` | `development.yml` | `127.0.0.1` | Bind address |
| `PORT` | `development.yml` | `8000` | Bind port |
| `LOG_LEVEL` | `development.yml` | `INFO` | Log level (`DEBUG`, `INFO`, `WARNING`, `ERROR`) |
| `CORS_ALLOW_ORIGINS` | `development.yml` | `["*"]` | Allowed CORS origins |
| `CORS_ALLOW_METHODS` | `development.yml` | `["*"]` | Allowed HTTP methods |
| `CORS_ALLOW_HEADERS` | `development.yml` | `["*"]` | Allowed HTTP headers |

## Design Notes

- **Single Starlette app** — `sse_app()` is the sole ASGI application; the health route and CORS middleware are injected directly onto it to prevent double-middleware stacking (which causes the `http.response.start` crash).
- **Separation of concerns** — Tools are thin wrappers around `OCRService`; business logic is independently testable.
- **Lifespan-managed client** — The Mistral client is initialized once at startup and shared across all tool calls.
- **Loguru logging** — Structured, colorized logs across all layers via Loguru.
- **Pydantic Settings** — Type-safe, `.env`-driven configuration with an LRU-cached singleton.