Case Studies / Custom MCP Server for Agent-Ready Fintech API Integration

Custom MCP Server for Agent-Ready Fintech API Integration

An MCP server built on top of a banking-infrastructure platform, exposing its APIs and documentation to Claude, Cursor and other AI tools so client developers can integrate without going through support. Around 42% fewer support tickets, with no changes to the underlying APIs.

• Custom MCP Server Development
• 2026
• Industry: FinTech

Initial Task

Description

The сlient is a banking-infrastructure company providing card issuance, payments, FX and Apple Pay / Google Pay. Banks, neobanks and fintech startups build their own products on top of its APIs, so how quickly their developers can integrate shapes the platform’s growth.

As the client base grew, those developers kept hitting the same integration questions: which endpoints to use, in what order, how to handle edge cases. The documentation answered all of it, but it sat outside the AI tools developers actually work in, so every question landed in the support queue.

The сlient approached Touchlane to make the platform usable directly inside those tools, under two constraints: no changes to the underlying APIs, and a clear line between answering a question and performing an operation.

Technologies

MCP OAuth2 Node.js Docker TypeScript

Challenges

01.

Describing a fintech API in a way models can act on

The platform's API surface was written for developers who read documentation and then write integration code. An AI agent needs the same capabilities described as tools it can discover, with clear inputs, outputs and failure modes. Card issuance, payments and FX each carry their own sequencing rules and edge cases, and those rules had to be expressed in the server itself rather than left for the model to infer.

02.

Permissions inside an AI tool

Guidance and action could not share the same level of access. Explaining how an endpoint works is safe to answer for anyone; calling it is not. The server had to carry a permission model with read-only access for guidance and guarded, authorized access for operations, with scoped access, validation and audit logging in line with the MCP specification's security requirements.

03.

No changes to the production APIs

The platform is live, with the Client's customers building on top of it. The MCP server had to be a dedicated integration layer on top of the existing architecture, with zero changes to the core codebase, and had to stay correct as the platform's own APIs and documentation evolve.

Workflow

01.

Product assessment and roadmap

Touchlane assessed the platform’s API surface and documentation, tracing the questions that reached support most often back to the endpoints behind them. That became the capability list the MCP server would expose, split between what an AI tool may explain and what it may execute.

02.

MCP server design

We designed the server as a dedicated integration layer over the existing architecture, with each capability described as a tool an AI client can discover and the documentation exposed alongside the APIs. The permission model was set here: read-only access for guidance, guarded access requiring explicit authorization for operations.

03.

Implementation

Our team built the server against the platform’s APIs with authentication, scoped access, input validation and audit logging, and no changes to the core codebase. Behaviour was verified in Claude and Cursor, the clients the Client’s developers actually use.

04.

Launch and enablement

The server was deployed to the Client’s own infrastructure with monitoring, and published to the MCP registry so AI clients can discover it. Handover included documentation and a Git repository with the full source code.

Overall Result

The same integration questions that used to go to support now resolve inside the AI tools the Client’s developers already work in. They self-serve, integrate independently and ship faster, and support ticket volume dropped by around 42%.

The platform is now reachable from any MCP-compatible client through a single server, built on top of the existing architecture with no changes to the underlying APIs and handed over with monitoring, documentation and full source code.