Sensible MCP. Extract structured JSON from PDFs, images, and files.

You're done copy-pasting data out of PDFs and invoices. This server takes messy, unstructured documents—whether they’re PDFs, images, or Word files—and converts them into clean, predictable JSON records using a robust parsing engine.

The system starts by letting you define what kind of document you're dealing with. You establish an entirely new category or type of document using create_document_type, and then you guide the specific data extraction for that type by creating a rule set with create_configuration. If you need to standardize your process, you can create a 'Golden' reference document via create_golden, which serves as the primary source for defining both data structure and quality control.

You manage these rules using functions like update_configuration or get_configuration, and if something goes wrong, you can delete configurations with delete_configuration or remove entire types with delete_document_type.

When you receive a document, the first step is figuring out what it is. You classify documents using classify_sync to get the type immediately, or you run an async job with classify_async if classification takes time. Once classified, you have several ways to extract data. If you've got a local file encoded as Base64, extract_sync pulls structured data instantly.

For files hosted online, you start a background process using generate_upload_url, then run the extraction with extract_from_url. You can even target multiple related documents at one spot by running extract_portfolio_from_url. If your async job needs specific rules, use generate_upload_url_with_config and execute the extraction via extract_from_url_with_config, or for immediate local pulls using a configuration ID, run extract_sync_with_config.

To keep track of all this background work, you can list past jobs with list_extractions. You'll also need to manage your reference materials; get_golden retrieves metadata on the current standard document, and you can inspect every text line and its coordinates from that master source using extract_text_from_golden. When you’ve pulled a batch of data, you don't want JSON blobs.

You compile multiple results into usable formats: generate_csv creates a standard CSV spreadsheet, while generate_excel builds a formatted Microsoft Excel workbook.

For large-scale operations, the server helps you manage access and history. Use get_auth_tokens to create temporary credentials for external reviewers accessing data securely. You can list all available configurations with list_configurations, or check historical versions and drafts of any rule set using list_configuration_versions. If you're auditing your process, get_extraction_statistics returns metrics on how much data has been extracted recently, and you can see every defined document type overview by calling list_document_types.

The system also lets you review all available reference documents for a specific type using list_goldens, or view the metadata about any given configuration with get_configuration_version.

The server handles everything from setup to final delivery. It allows you to pull data instantly on local files, run complex background jobs against remote URLs, and aggregate those results into spreadsheets ready for your team to use.