ETL-D MCP Server 🌊

A deterministic data middleware for AI agents, exposed as a standard Model Context Protocol (MCP) server.

The Problem: The "Data Tax" in LLMs

Asking an LLM to parse a 10,000-row CSV, a Spanish Norma 43 bank statement, or an EDI 850 file directly in the context window leads to three major issues:

Token Exhaustion: Sending raw B2B formats to a context window is highly inefficient.
Hallucinations: LLMs struggle with strict spatial alignment and counting. A single misplaced comma in a financial trade history is catastrophic.
Non-Determinism: You cannot build reliable B2B pipelines if the extraction format changes based on the LLM's "mood."

The Solution: Waterfall Architecture

ETL-D stops agents from attempting to "read" raw B2B data. Instead, the agent routes the file/string to this MCP server, which processes the data using a strict 3-Layer Waterfall Architecture:

L1 - Heuristic (0-Shot): Standard Python parsers (regex, dateutil, strict structural rules). If the pattern is known, it parses instantly with 0% hallucination risk.
L2 - Semantic Routing: Maps obfuscated headers to strict Pydantic schemas using embedding-based alignment.
L3 - LLM Fallback: Only triggered for high-entropy "free-text" noise.

The agent receives a perfectly structured, flattened JSON array ready for reasoning.

🚀 Quick Start (Claude Desktop)

To use ETL-D with Claude Desktop, add the following to your claude_desktop_config.json:

{
  "mcpServers": {
    "etld": {
      "command": "npx",
      "args": [
        "-y",
        "@pablixnieto2/etld-mcp-server@3.2.2"
      ],
      "env": {
        "ETLD_API_KEY": "your_api_key_here"
      }
    }
  }
}

Authentication & API Key

ETL-D uses a hybrid compute engine (Python backend) to run the deterministic parsing layers. To prevent abuse, it requires an API Key.

You can provision an API key by getting Prepaid Credits or a Monthly Subscription here: 👉 Get ETL-D API Key (Requires Stripe/Unkey auth)

🛠️ Available MCP Tools

When connected, your agent will automatically have access to the following deterministic pipelines:

1. B2B & Financial Parsing

parse_bank_statement: Transforms Spanish Norma 43 (N43) raw files into structured JSON (account info, balances, multi-line transactions).
parse_edi: Parses raw ANSI X12 EDI strings (optimized for 850 POs) into structured JSON.
parse_trade_history: Deterministic extraction of complex broker trade exports.
generate_sepa_xml: Converts JSON into SEPA Direct Debit (PAIN.008.001.02) XML files.

2. Document Intelligence

pdf_to_spatial_markdown: Converts raw PDFs into LLM-optimized Markdown, preserving table structures and spatial context before the LLM attempts to read it.
extract_invoice / extract_resume: Strict schema extraction for standard business documents.

3. Core Enrichment (Atomic Tools)

enrich_amount: Extracts precise financial data (float values, ISO currency codes) from messy strings (e.g., "Total: 1.240,50€").
enrich_date: Resolves human-readable dates (e.g., "next Tuesday at 5pm") with timezone awareness into ISO 8601.
enrich_address: Transforms messy global address strings into standardized components.
accounting_map: Omnidirectional mapper that translates unstructured expense concepts into standard accounting frameworks (e.g., US GAAP, ES PGC).

🏗️ Architecture & Ecosystem

This repository (etld-mcp-server) contains the lightweight TypeScript/Node.js MCP Server. It acts as a stateless bridge between your AI Agent (like Claude Desktop) and the ETL-D Cloud Engine.

The ETL-D Ecosystem:

MCP Server (This repo): The native integration for Claude and other MCP-compatible agents.
Cloud API: The heavy lifting (Python/FastAPI engine) runs securely on api.etl-d.net.
Python SDK: Building your own custom agent in Python? You don't need MCP. Just use our native Python SDK: pip install etld.

License

MIT

etld-mcp-server

Quick Install