Financial Operations Agent

An agentic financial operations assistant built with the Claude Agent SDK, exposed via an Express HTTP API. Designed for a large-scale digital bank serving millions of customers across checking accounts, savings accounts, and digital wallets.

Architecture

Express Server (:3000)
  POST /api/chat   → Claude Agent (claude-sonnet-4-5)
  GET  /api/health → Health check
                          │
                          ▼
                    5 Custom MCP Tools
                          │
                          ▼
                    Mock Data Layer
              (accounts, transactions, risk profiles)

The agent runs as a single Claude instance with custom tools registered via createSdkMcpServer. It connects to Azure AI Foundry as the model provider.

Tools

Tool	Description
get_account_balance	Look up account balance, status, and metadata
get_transaction_history	Query transactions with date range and limit filters
calculate_risk_score	Get risk score (0-100), risk factors, and recommendation
generate_compliance_report	Generate AML, KYC, or daily summary reports
flag_suspicious_activity	Flag a transaction and create a fraud investigation case

Setup

Prerequisites

• Bun runtime
• Claude Code CLI installed (claude command available)
• Azure AI Foundry credentials (or a direct Anthropic API key)

Install

bun install

Configure

Copy the example env file and fill in your credentials:

cp .env.example .env

ANTHROPIC_FOUNDRY_API_KEY=your-azure-foundry-api-key
ANTHROPIC_FOUNDRY_BASE_URL=https://your-resource.services.ai.azure.com
PORT=3000

To use a direct Anthropic API key instead of Azure Foundry, set ANTHROPIC_API_KEY and remove the Foundry variables. You'll need to update src/agent.ts to skip the initAzureFoundry() call.

Run

bun run start

Or with auto-reload:

bun run dev

API

GET /api/health

Returns server status.

curl http://localhost:3000/api/health

{ "status": "ok", "timestamp": "2026-03-15T13:35:05.027Z" }

POST /api/chat

Send a message to the agent.

curl -X POST http://localhost:3000/api/chat \
  -H "Content-Type: application/json" \
  -d '{"message": "Check the balance and risk score for account ACC-004"}'

{ "response": "Account ACC-004 belongs to Can Öztürk..." }

Example queries

# Account investigation
curl -X POST http://localhost:3000/api/chat \
  -H "Content-Type: application/json" \
  -d '{"message": "Check the balance and risk score for account ACC-004"}'

# Compliance report
curl -X POST http://localhost:3000/api/chat \
  -H "Content-Type: application/json" \
  -d '{"message": "Generate an AML compliance report"}'

# Fraud flagging
curl -X POST http://localhost:3000/api/chat \
  -H "Content-Type: application/json" \
  -d '{"message": "Flag transaction TXN-030 — appears to be structuring to avoid reporting thresholds"}'

# Transaction analysis
curl -X POST http://localhost:3000/api/chat \
  -H "Content-Type: application/json" \
  -d '{"message": "Show me the recent transactions for account ACC-008 and assess if anything looks suspicious"}'

Project Structure

src/
├── index.ts              # Express server
├── agent.ts              # Agent config, system prompt, Azure Foundry init
├── tools/
│   ├── index.ts          # Barrel export
│   ├── helpers.ts        # toolResponse/toolError utilities
│   ├── balance.ts        # get_account_balance
│   ├── transactions.ts   # get_transaction_history
│   ├── risk-score.ts     # calculate_risk_score
│   ├── compliance.ts     # generate_compliance_report
│   └── fraud.ts          # flag_suspicious_activity
└── data/
    ├── accounts.ts       # 10 mock accounts
    ├── transactions.ts   # 31 mock transactions
    └── risk-profiles.ts  # 10 mock risk profiles

Mock Data

The agent operates on realistic mock data representing a Turkish digital bank:

• 10 accounts across checking, savings, and digital wallet types
• 31 transactions including normal activity, suspicious patterns, and structuring attempts
• 10 risk profiles ranging from low (score 5) to critical (score 92)

Notable accounts for testing:

• ACC-004 — Frozen account, risk score 92, multiple large withdrawals
• ACC-008 — Active account, risk score 68, crypto structuring pattern
• ACC-001 — Normal active account, low risk

Tech Stack

• Runtime: Bun
• Agent: @anthropic-ai/claude-agent-sdk
• Server: Express 5
• Validation: Zod 4
• Model Provider: Azure AI Foundry (claude-sonnet-4-5)