Model Context Protocol (MCP)

Definition. An open protocol that standardises how AI applications connect to external data sources and tools, separating the host application, the client and the server that exposes resources, prompts and tools. It replaces bespoke per-integration glue with a common transport and capability-negotiation layer, so one server implementation can serve many hosts and one host can consume many servers.

How it works. Run or connect to an MCP server that declares its capabilities. The host discovers available tools and resources, then exposes them to the model as callable functions. Configure authentication at the server, scope permissions narrowly, and pin server versions. For internal systems, write a thin server that wraps your existing API rather than granting the model direct database access.

Example. An operations team needs its assistant to read from Jira, query a Postgres warehouse and post to Slack. Instead of three custom integrations, they deploy three MCP servers. The assistant host connects to all three, discovers eleven tools, and the model composes them: query the warehouse for churn by cohort, open a Jira ticket with the finding, and post a summary to a Slack channel. Adding a fourth system later requires only a new server, not a change to the host.

Use case. MCP is the right choice when you are building a platform that must integrate with many systems, or when you want a tool integration to be reusable across several AI clients. It also gives you a clean security boundary, because each server can be permissioned and audited independently.

Warnings. Every MCP server you connect is code running with your credentials. Audit third-party servers before use, restrict their scopes, and never connect a server that can write to production systems without a confirmation layer. Tool descriptions from a server are model input and can carry injection payloads.

Limitations. The protocol is young and implementations vary in completeness. Latency adds up when a request fans out across several servers. Discovery can flood the context with dozens of tool definitions, hurting selection accuracy, so most hosts need tool filtering or lazy loading.

Pros and cons
ProsCons
Standardizes how applications expose tools and data to models.Implementations vary and requests across many servers can add latency.
← Browse all glossary terms