I built an MCP (Model Context Protocol) server that gives AI assistants structured, real access to financial product data — starting with business loans.
By “replace comparison sites,” I mean enabling AI assistants to access the same underlying product data directly, instead of sending users through Google → SEO pages → lead forms.
What bothered me: People already ask ChatGPT and Claude questions like “What’s the best business loan for X?” But today the models either hallucinate rates, recommend lenders that don’t offer those products, or give generic advice scraped from blogs.
I noticed users copy-pasting results from loan comparison sites into ChatGPT just to sanity-check them — which felt backwards.
What this does instead: With the MCP server installed, you can ask something like: “Compare SBA loans for a $500k equipment purchase with a 720 credit score.”
The AI queries structured lender data and returns actual offers — rates, terms, and estimated monthly payments — conversationally.
Technical notes (things I learned the hard way):
- MCP server in Node.js / TypeScript running on ECS Fargate
- Pre-aggregated lender data in DynamoDB (real-time lender APIs were a dead end)
- 20+ granular, strongly-typed tools instead of a generic “search” endpoint
- HTTP streamable transport
- Built SOC 2-compliant from day one
Business model (briefly): Free for borrowers. Lenders pay standard lead-gen fees. The same MCP infrastructure also feeds our B2B loan-origination software.
Distribution so far:
- Live on Smithery.ai and the Cursor MCP directory
- Submitted to the Claude MCP directory
- Submitted to the ChatGPT app store
- One-click Claude Desktop installer: https://extensions.securelend.ai
My bet is that discovery moves from pages to conversations, and MCP-style servers become the integration layer that makes that possible.
GitHub: https://github.com/SecureLend/mcp-financial-services
Docs: https://docs.securelend.ai
Would love feedback on:
- MCP server architecture and tool design
- Handling compliance in AI-native fintech
- Whether AI-native distribution is actually viable
Happy to answer technical or product questions.