Finance Reconciliation

MCP Server for Month-End Close Reconciliation
Finance Shared Services · GBS/SSC · Auxiliobits

Automate the most manual, error-prone activities in a shared services team:
bank reconciliation, GL close, intercompany matching, and month-end close tracking.

---

What it does

MCP Tool	Description
match_bank_statement	Match bank CSV vs GL cash ledger — ±3 day date proximity, exact amount
reconcile_gl_accounts	Compare prior vs current GL extract — new entries, reversals, variances
check_intercompany_balances	Entity A vs Entity B — gross mismatch, FX component, risk flag
classify_reconciliation_breaks	Rules-based break classification with recommended next actions
track_month_end_close	12-task close tracker with overdue detection and risk rating

---

Quick Start

1. Install

git clone https://github.com/auxiliobits/auxilab-mcp-finance-recon
cd auxilab-mcp-finance-recon
pip install -e ".[dev]"

2. Generate test data

python data/generate_test_data.py

This creates data/sample_data/ with 7 synthetic CSV files — no real data required.

3. Run the standalone demo

python demo.py

Expected output: full 5-step reconciliation flow ending with a Month-End Readiness Report.

4. Start the MCP server (for Claude Desktop / agent use)

python src/server.py

---

MCP Server Config (claude_desktop_config.json)

{
  "mcpServers": {
    "finance-recon": {
      "command": "python",
      "args": ["/path/to/auxilab-mcp-finance-recon/src/server.py"]
    }
  }
}

---

Tool Reference

match_bank_statement

{
  "bank_csv": "<CSV string>",
  "gl_csv": "<CSV string>",
  "opening_balance": 500000.00
}

Returns: reconciliation_summary, matched_pairs, unmatched_bank_items, unmatched_gl_items

Matching logic:

• Amount match: exact within ±$0.01
• Date proximity: ±3 calendar days
• Tiebreaker: rapidfuzz token_sort_ratio on description

---

reconcile_gl_accounts

{
  "period_a_csv": "<prior period CSV>",
  "period_b_csv": "<current period CSV>",
  "account_name": "Accrued Liabilities — 2110"
}

Returns: reconciliation_statement, new_entries, removed_entries, reversals, common_entries

---

check_intercompany_balances

{
  "entity_a_csv": "<Entity A ledger CSV>",
  "entity_b_csv": "<Entity B ledger CSV>",
  "entity_a_name": "Auxiliobits India",
  "entity_b_name": "Auxiliobits Singapore",
  "base_currency": "USD"
}

Returns: summary (with mismatch_risk: NONE/LOW/MEDIUM/HIGH/CRITICAL), matched_transactions, entity-only lists, mismatch_detail

Demo scenario: Entity A = $125,000 receivable · Entity B = $123,500 payable → $1,500 mismatch, flagged HIGH

---

classify_reconciliation_breaks

{
  "breaks_csv": "<unreconciled items CSV>"
}

CSV columns: reference, date, description, amount, source

Classification categories:

Category	Rule
Timing Difference	Keywords: transit, outstanding, cut-off, clearing
Missing Entry	Keywords: missing, not posted, unposted
Duplicate Posting	Keywords: duplicate, double post, twice
Currency Rounding	Amount ≤ $10, or FX/forex keywords
Intercompany Mismatch	Keywords: interco, entity, subsidiary
Unknown	No rule matched — escalate

---

track_month_end_close

{
  "tasks_csv": "<tasks CSV>",
  "close_deadline": "2024-11-30",
  "as_of_date": "2024-11-29"
}

CSV columns: account_name, owner, status, due_date

Valid statuses: complete, in progress, not started, overdue, on hold

Risk logic:

• CRITICAL: deadline ≤ 1 day with outstanding tasks, or deadline passed
• HIGH: deadline ≤ 2 days with >3 outstanding, or ≥3 overdue tasks
• MEDIUM: completion < 75% or any overdue
• LOW: on track

---

Project Structure

auxilab-mcp-finance-recon/
├── src/
│   ├── server.py               # MCP server — tool registry + dispatch
│   └── tools/
│       ├── __init__.py
│       ├── bank_matcher.py     # Bank ↔ GL matching (Pandas + rapidfuzz)
│       ├── gl_reconciler.py    # Prior vs current GL extract diff
│       ├── intercompany_checker.py  # Entity A vs Entity B
│       ├── break_classifier.py # Rules-based break classifier
│       └── close_tracker.py   # Month-end task tracker
├── data/
│   ├── generate_test_data.py  # Synthetic CSV generator
│   └── sample_data/           # Generated CSVs (gitignored)
├── tests/
│   └── test_tools.py
├── demo.py                    # Full 5-step demo runner
├── pyproject.toml
└── README.md

---

Tech Stack

Component	Library
MCP protocol	mcp Python SDK >= 1.0.0
Matching logic	pandas >= 2.0, numpy
Fuzzy description matching	rapidfuzz (falls back to difflib)
CSV parsing	pandas.read_csv
Break classification	Rules-based (regex) — no LLM required

---

Demo Data Summary

File	Rows	Notes
bank_statement.csv	50	Unique amounts, full month
gl_cash_ledger.csv	48	45 match bank (±1-2 day drift), 3 GL-only timing items
gl_prior_period.csv	20	October GL extract
gl_current_period.csv	25	15 carry-forward + 8 new + 2 reversals
entity_a_ledger.csv	3	$125,000 receivable
entity_b_ledger.csv	3	$123,500 payable — $1,500 mismatch
month_end_tasks.csv	12	4 complete, 3 overdue, deadline in 1 day → HIGH risk
breaks_sample.csv	8	Covers all 6 break categories

---

License

MIT · Built for the Auxiliobits Hackathon — Financial Reconciliation Pillar