Docstats: Readability Metrics API and MCP Server

Docstats is a versatile Python application that calculates a comprehensive set of readability metrics for textual content. It can be run as a standard HTTP API (using FastAPI) or as an MCP (Model Context Protocol) server in two modes: STDIO or Streamable HTTP.

The application can process text from three types of sources:

1. Direct text input.

2. A publicly accessible web URL (supports both HTML content and PDF files).

3. A Google Cloud Storage (GCS) URI pointing to a PDF file.

Features

• Calculates numerous readability scores (Flesch Reading Ease, Flesch-Kincaid Grade, Gunning Fog, etc.).

• Provides basic text statistics: syllable count, word count, sentence count.

• Accepts input as direct text, web URL (HTML/PDF), or GCS PDF URI.

• HTML parsing for web URLs to extract relevant content using BeautifulSoup.

• PDF text extraction from web URLs and GCS using pypdf.

• Multi-mode operation: FastAPI HTTP server, MCP STDIO server, or MCP Streamable HTTP server.

Related MCP server: crawl-mcp-server

Setup

1. Clone the repository (if applicable).

2. Create and activate a virtual environment using uv (recommended):

   python -m venv .venv # Or your preferred venv creation method
   source .venv/bin/activate  # On Windows: .venv\Scripts\activate

3. Install dependencies: If you have requirements.txt:

   uv pip install -r requirements.txt

   If you only have pyproject.toml or want to ensure all dependencies from it are installed (recommended):

   uv pip install -e . # Installs the project in editable mode with its dependencies
   # or for just dependencies if not installing the project itself:
   # uv pip install . 

   Ensure your pyproject.toml lists all necessary dependencies like fastapi, uvicorn[standard], textstat, py-readability-metrics, google-cloud-storage, pypdf2, beautifulsoup4, mcp>=1.9.3, starlette, anyio.

4. Google Cloud Authentication (for GCS PDF processing): Ensure you have Application Default Credentials (ADC) configured in your environment. This typically involves running:

   gcloud auth application-default login

   Refer to Google Cloud ADC documentation for more details.

Running the Application

The main.py script uses command-line flags to determine how it starts.

1. As a FastAPI HTTP Server

This mode provides a traditional RESTful API. For development with auto-reload, it's best to run Uvicorn directly pointing to the fastapi_app module.

Recommended Development Startup (with auto-reload):

uv run uvicorn fastapi_app:fastapi_app --host 127.0.0.1 --port 8000 --reload

• Adjust --host and --port as needed.

• API documentation (Swagger UI) will be available at http://127.0.0.1:8000/docs.

Alternative Startup (using the main.py entry point):

uv run python main.py --server-type fastapi --host 127.0.0.1 --port 8000

• Note: Uvicorn's --reload feature works best when Uvicorn imports the app string (main:fastapi_app) directly.

Example API Call (using cURL for a web URL):

curl -X POST "http://127.0.0.1:8000/scores/" -H "Content-Type: application/json" -d '{ "web_url": "https://www.example.com" }'

Example API Call (using cURL for direct text):

curl -X POST "http://127.0.0.1:8000/scores/" -H "Content-Type: application/json" -d '{ "text": "This is a sample text for readability analysis." }'

2. As an MCP STDIO Server

This mode runs the MCP server communicating over standard input/output. This is useful for direct integration with other processes or MCP clients that use STDIO (e.g., an AI Agent CLI).

Command using python main.py:

uv run python main.py --server-type mcp

• This will start the mcp_generic_app instance defined in main.py using the STDIO transport.

Using with an MCP Client (e.g., AI Application / Gemini CLI / Claude): If you have an MCP client configured (like Gemini CLI or Claude Code), you can define this server in its settings.

For example, in a ~/.claude/settings.json or ~/.gemini/settings.json (or similar client configuration file for your agent):

{
  "mcpServers": {
    "readability_docstats": {
      "command": "uv",
      "args": ["run", "python", "/PATH/TO/DOCSTATS/main.py", "--server-type", "mcp"],
      "cwd": "/PATH/TO/DOCSTATS"
    }
  }
}

Replace /PATH/TO/docstats with the correct absolute path to this project directory if your MCP client runs from a different context. Once configured, you could invoke the tool via the client, e.g.: @readability_docstats get_readability_scores text="Some sample text."

3. As an MCP Streamable HTTP Server

This mode runs the MCP server over HTTP.

Command using python main.py:

uv run python main.py --server-type mcp-http --host 127.0.0.1 --port 8001

• Adjust --host and --port as needed. A different port (e.g., 8001) is recommended to avoid conflict with the FastAPI server.

• The MCP endpoint will typically be available at http://127.0.0.1:8001/mcp.

• To get plain JSON responses instead of SSE streams (useful for some clients or debugging):

  uv run python main.py --server-type mcp-http --mcp-http-json-response --port 8001

Development and Testing

Testing

The project includes a comprehensive test suite using pytest.

Test Types

1. Unit Tests (test_unit.py): Uses mocks for httpx, google-cloud-storage, and pypdf. These are fast and do not require internet or cloud credentials. They cover core extraction logic and MCP tool execution.

2. Integration Tests (test_main.py): Uses FastAPI's TestClient to verify end-to-end flow. Some tests are marked as slow because they require network access to fetch live web pages and GCS PDFs.

Test Coverage

• Direct Text Input: Verifies score calculation for both short and medium texts.

• Web & PDF Extraction: Tests HTML and PDF extraction from both live URLs and mocked responses.

• GCS PDF Extraction: Verifies GCS integration (mocked and live).

• MCP Tools: Verifies that the Model Context Protocol tools return correctly formatted JSON-RPC responses.

• Validation & Errors: Ensures 422 errors are returned for invalid inputs or source conflicts.

Running Tests

Run all tests:

uv run pytest

Run only unit tests (fast, no network):

uv run pytest test_unit.py

Run tests excluding slow integration tests:

uv run pytest -m "not slow"

Benchmarking & Samples

The project includes a Readability Golden Set in the samples/ directory, containing texts of varying complexity (Primary, Middle, Academic, Legal).

Baseline Analysis

You can run a baseline analysis to see how the engine scores these different levels:

uv run python baseline_analysis.py

This script:

1. Processes all files in samples/.

2. Prints a summary table of grade standards and word counts.

3. Saves the full metric breakdown to samples/baseline_results.json.

Note: The samples are used to verify that code or logic changes (e.g., extraction or normalization) don't unintentionally shift the readability scores of known content.

Dependencies

Manage Python dependencies using uv and pyproject.toml.

• To add a new dependency: uv add <package>

• To sync environment: uv sync

Available Readability Scores

The API returns the following scores (where applicable):

• Flesch Reading Ease

• Flesch-Kincaid Grade Level

• Gunning Fog Index

• SMOG Index

• Automated Readability Index (ARI)

• Coleman-Liau Index

• Linsear Write Formula

• Dale-Chall Readability Score

• Spache Readability Score (Note: May return null if text doesn't meet specific criteria, e.g., word count)

• Text Standard (Consensus grade level)

• Syllable Count

• Word Count

• Sentence Count

License

Apache 2.0

Disclaimer

This is not an officially supported Google product.