Invoice & Receipt to JSON + Excel, Totals Checked (MCP server)

A deterministic invoice and receipt reader for AI agents, served as an MCP server. There is no LLM inside, so the same file gives the same output every time. It runs on CPU and does not call any paid API.

Use it for: accounts-payable intake, expense receipts, bookkeeping imports, matching invoices to purchase orders or bank lines, and checking supplier invoices before they are paid.

Why agents pick it: an agent that books a wrong total costs real money. This tool does not just read the figures, it does the arithmetic: every line's qty x unit price is recomputed, line items are summed against the net, and net + VAT is checked against the gross. Your agent gets a plain all_ok flag and a list of the checks, so it knows when to trust the result and when to ask a human.

Tool

What your agent gets back

invoice_to_json

supplier, invoice_number, invoice_date and due_date (ISO), vat_number / vat_numbers, currency, net, vat, gross, discount, shipping, and line_items (description, qty, unit_price, discount_pct, vat_rate, tax_amount, amount). checks lists every reconciliation made, with the computed and stated figure. all_ok is true only when every check that could be made passed. Plus an XLSX with Invoice, Line items and Checks sheets.

Reads PDFs with a text layer directly, and OCRs scanned PDFs and photos (PNG, JPG, TIFF) with Tesseract. Labels are recognised in English, with common German, French and Dutch equivalents (Rechnungsnr., Total HT, TVA, Totaal, BTW, MwSt). Amounts in 1,234.56, 1.234,56 and 3,89 € styles are all handled.

Connect

MCP endpoint (streamable HTTP):

https://doubleaigents--invoice-receipt-to-json-mcp.apify.actor/mcp
Authorization: Bearer <APIFY_TOKEN>

Claude Desktop / Cursor style config:

{ "mcpServers": { "invoices": {
    "url": "https://doubleaigents--invoice-receipt-to-json-mcp.apify.actor/mcp",
    "headers": { "Authorization": "Bearer <APIFY_TOKEN>" } } } }

Tool inputs

invoice_to_json(file_url | file_base64, filename?, lang="eng")

• Send exactly one of file_url (public http/https) or file_base64. Supported files are PDF, PNG, JPG, TIFF, BMP and WEBP, up to 50 MB and 500 pages.
• The XLSX comes back as a key-value store URL. Files under 8 MB also come back inline as base64.

Pricing (pay per event)

Event	USD
page-processed: each page of the input file, after extraction and reconciliation	0.03

You are charged only after a call completes. If your run's maximum charge cannot cover the page count, the call is refused before any work is done.

Limits worth knowing

• It is a layout reader, not a language model. It finds fields by their labels and position. Unusual layouts can leave a field empty; when that happens it is left null, never guessed.
• all_ok: false means a check failed or there was nothing to check (no total found). Read problems.
• The supplier is taken from a "From/Supplier/Sold by" block, else the letterhead. On invoices where the customer's name is printed first, check this field.
• Handwriting is not supported. Poor photos lower OCR accuracy; the reconciliation is what flags it.

Batch mode

You can also run the Actor normally with fileUrl (or a list in fileUrls). It writes one dataset item per file with the fields, line items, checks and a link to the XLSX.

Who runs it

Double AIgents Ltd, UK (company no. 17513072). Your input files are processed for the call. Outputs sit in the run's Apify storage, which Apify removes under its data-retention policy. Need volume pricing, a private deployment or a custom document pipeline? https://www.doubleaigents.com

More tools from Double AIgents

• PDF Tables to Excel + True PII Redaction — tables from PDFs and scans to XLSX with every figure double-read, and true redaction of personal data with a verification log.
• Bank Statement PDF to CSV/Excel — every transaction, with running balances and opening/closing totals verified.
• PDF Compare — every added, removed, changed or moved passage between two PDF versions, plus an annotated diff PDF.