Model Context Protocol: hosts, clients and servers
What MCP standardises, how its three roles fit together, and where it sits in a finance flow where an agent reads accounts or drafts payments.
The Model Context Protocol (MCP) is an open protocol for connecting applications built on language models to external data and tools. The current specification is versioned by date; the one read for this episode is 2026-07-28.
MCP uses JSON-RPC messages between three roles:
- Hosts, the AI applications that start connections;
- Clients, connectors inside the host;
- Servers, services that provide context and capabilities.
Servers can offer resources (context and data), prompts (templated messages) and tools (functions the model may ask to run). Clients can offer elicitation, a way for a server to ask the user for more information. Optional extensions, such as tasks for long-running work, must be negotiated by both sides.
In a finance flow, a bank or fintech might run an MCP server that exposes read-only account data as resources and a payment-draft function as a tool. The agent's host decides which of these the model sees.
Limits. MCP moves requests and results; it does not decide whether an action is authorised. The specification's own security principles say users must explicitly consent to data access and to tool invocation, and that tool descriptions should be treated as untrusted unless they come from a trusted server. Mandates, payment limits and approvals sit outside the protocol and must be built around it.
Sources: Model Context Protocol specification (version 2026-07-28), Model Context Protocol specification: Architecture