Recipe

Build an application assistant with MCP

Compose the public MCP tools into an assistant: eligibility_criteria to qualify, get_quote to price, how_to_apply to guide, then hand the user to the hosted application — every figure live and reconciled.

2 min read

4 toolsChained

Live figuresNo hallucination

Hand-offTo apply

The flow

A good lending assistant follows the same order a human adviser would. Register the public MCP server as a tool source, then let the model call, in sequence: eligibility_criteria to confirm the company can apply, get_quote to show a real figure, how_to_apply to explain the steps, and list_products if the user is choosing between products.

Why MCP beats scraping

Because the tools read the hub's authoritative product and pricing config, the assistant quotes the real numbers rather than stale figures scraped from a page — the same numbers the marketing site shows. That is the whole point of the MCP server: accurate, tool-callable Credicorp data with no hallucination.

Hand off cleanly

The assistant should never pretend to take the application itself — that runs on the hosted journey. End the conversation by handing the qualified user to apply online. If your assistant is authenticated as a partner and needs to do more than read, move to the partner MCP server.

Frequently asked questions

Can the assistant submit the application?

Not via the public MCP tools — they are read-only. The assistant qualifies, quotes and guides, then hands the user to the hosted apply journey. Submitting programmatically requires the token-gated partner ring.

How do I stop the model inventing figures?

Give it the MCP tools and instruct it to quote only from tool results. Because the tools return the hub's authoritative figures, a tool-grounded answer is accurate; a model answering from memory is not.