ETL-D MCP Server 🌊

ETL-D is a deterministic data middleware designed to act as a shield for AI Agents. It stops LLMs from "hallucinating" over structured data by providing a strict, 3-layer parsing architecture via the Model Context Protocol (MCP).

⚠️ The Problem: The "Data Tax" & Hallucinations

Standard LLMs are terrible at reading raw B2B files (CSV, PDF, EDI, Norma 43). They suffer from:

1. Token Exhaustion: Sending a 5,000-row CSV to context is a waste of money.

2. Precision Loss: LLMs struggle with spatial alignment. A misplaced comma in a bank statement is a financial catastrophe.

3. Non-Determinism: You can't build a reliable pipeline if the output format depends on the LLM's "mood."

Related MCP server: document-parser

✅ The Solution: 3-Layer Waterfall Architecture

The ETL-D engine processes every request through a strict hierarchy:

• Layer 1 (Heuristic): 100% Python-native deterministic parsers. 0% Hallucination risk. ~70ms latency.

• Layer 2 (Semantic): Column-to-Schema alignment using embedding-based routing.

• Layer 3 (LLM Shield): Strict JSON enforcement using Llama 3.3 70B with additionalProperties: false constraints for high-entropy noise.

🚀 Quick Start (Claude Desktop)

1. Open your claude_desktop_config.json.

2. Add the ETL-D server (Ensure you use version 3.3.3):

{
  "mcpServers": {
    "etld": {
      "command": "npx",
      "args": [
        "-y",
        "@pablixnieto2/etld-mcp-server@3.3.3"
      ],
      "env": {
        "ETLD_API_KEY": "YOUR_API_KEY_HERE"
      }
    }
  }
}

🔑 How to Get an API Key?

• Option A: Manual Purchase (Human) Get a Starter Pack ($5 / 5,000 credits) or a Subscription at api.etl-d.net.

• Option B: Zero-Touch Provisioning (Agent-Led) 🤖 If you don't have a key, simply ask Claude: "I don't have an ETL-D key, can you help me get one?". The agent will call the /provision tool, generate a Stripe Checkout link for you, and automatically set up the key once paid. Zero-touch, human-in-the-loop.

🛠️ Available MCP Tools

1. Financial & B2B Heavy Lifting parse_bank_statement: Support for Spanish Norma 43 (N43). Turns raw bank files into clean JSON.

parse_trade_history: Deterministic extraction of trades, fees, and dividends from complex broker exports.

parse_edi: ANSI X12 EDI parser (Optimized for 850 Purchase Orders).

generate_sepa_xml: JSON to PAIN.008 (Direct Debit) XML generator.

2. Document Intelligence pdf_to_spatial_markdown: Crucial for Agents. Converts PDFs to Markdown preserving table structures before the LLM reads them.

extract_invoice / extract_resume: High-accuracy schema extraction for standard B2B documents.

3. Atomic Enrichment (1 Credit/call) enrich_amount: Cleans "Total: 1.240,50€" into {amount: 1240.50, currency: "EUR"}.

enrich_date: Resolves human-readable dates ("next Friday at 5pm") with Timezone awareness.

enrich_address: Standardizes global messy addresses into structured components.

accounting_map: Maps concepts to ES PGC, US GAAP, or IFRS.

🏗️ Ecosystem

Cloud Engine: Hosted at api.etl-d.net (Python/FastAPI).

Python SDK: pip install etld.

n8n Nodes: Available in the n8n community as n8n-nodes-etld.

⚖️ License

MIT - Created by Pablixnieto2