Using the Model Context Protocol (MCP) in Financial Services

In November 2024, Anthropic open-sourced the Model Context Protocol (MCP), a standard for connecting AI assistants to external data sources and tools. For financial services, MCP offers a single protocol to expose internal APIs, databases, and documents to Claude (and other MCP clients) without building one-off integrations per system.

What MCP is and why it matters

Today, every data source or tool often requires a custom connector. MCP defines a protocol so that:

• MCP servers expose resources (e.g. tables, files, tools) in a standard way.
• MCP clients (e.g. Claude Desktop, or your own app) connect to those servers and let the model read data and invoke tools through a common interface.

Anthropic ships an MCP specification, SDKs, and support in Claude Desktop (including Claude for Work). Pre-built or community MCP servers exist for systems like Postgres, Google Drive, Slack, and GitHub.

For banks and insurers, the value is building or adopting MCP servers for core systems (e.g. product config, risk data, document stores) so that one assistant can work across them without N custom integrations.

MCP on Databricks

Databricks Agent Framework supports MCP natively. You can register MCP servers as tools for your agents, and Agent Bricks Multi-Agent Supervisor can orchestrate agents that connect to multiple MCP servers. Unity Catalog governs which MCP tools each agent can access, maintaining the security boundary.

How Claude uses MCP

Claude is well-suited to implementing and using MCP: you can build or configure MCP servers that expose read-only or controlled-write tools, and Claude Desktop (or a custom client) connects to them. Access can be scoped so the model only sees what each user or role is allowed to see. That is important for internal data: you define the server and its permissions; the model does not get raw DB credentials.

Use cases in finance

• Internal Q&A: An MCP server over your policy or product knowledge base; analysts ask questions in natural language and get answers grounded in that data.
• Controlled tool use: Servers that expose "get eligibility," "fetch quote," or "log action" as tools, with strict validation and logging on your backend.
• Unified assistant: One Claude (or other MCP client) session that can query multiple MCP servers (risk, documents, CRM) so users do not switch between tools.

All of this stays within your control: the model sends requests to your MCP servers; you decide what is exposed and how it is logged.

Building or adopting MCP servers

You can implement an MCP server in your stack (Anthropic provides SDKs and docs) that wraps your APIs or databases. Alternatively, you use or adapt existing servers and add ones for proprietary systems. For production, you will want authentication, network isolation, and audit logging so MCP does not become a back door. The protocol is open, so other clients (IDEs, agents, internal apps) can reuse the same servers once they are in place.