DEEP DIVE

Model Context Protocol (MCP)

How agents connect to your core banking systems, document repositories, and APIs.

🔌 What MCP Actually Solves

Before MCP, connecting an agent to a new tool meant custom integration work for every combination of model and system. MCP standardizes that connection — an agent that speaks MCP can connect to any MCP-compliant server (a database, an API, a document store) without bespoke glue code for each one.

🏦 Finance Scenarios

🏛️

Core Banking Read Access

An MCP connector that lets a Chat Agent query (never write) account balances or transaction history to answer a status question — scoped to read-only permissions.

📊

Market/Reference Data

Connecting to a public reference-rate or FX data source so an agent can cite a current rate rather than an outdated one baked into a document.

📁

Document Repository

Connecting to a document management system so an agent's Knowledge scope always reflects the current, approved version of a policy — not a stale copy.

🛡️ Permissions Hygiene for MCP in Finance

Match agent rights to MCP rights — if the agent should never write data, the MCP connector's credentials should be read-only at the system level, not just "asked nicely" in the prompt.
Scope narrowly — connect to the specific table/endpoint needed, not a broad general-purpose connector, to keep the blast radius small if something misfires.
Log every call — MCP tool calls should be as auditable as the agent's chat output; this is what makes "numbers traceable to tool calls" (Module 5) actually true.

A live system connection raises the stakes. Once an agent can query a real core banking system, treat its access review with the same rigor as any other system integration — this is a security control, not just a course exercise.