Model Context Protocol is a specification for how an AI agent finds out what tools are available to it and invokes them. In API terms, an MCP server is a façade over capabilities — each tool has a name, a description, an input schema, and an implementation that usually calls your existing API underneath.
Whether you need one depends on whether you want agents to be able to use what you offer. If your answer is yes, then the honest framing is that MCP is a new consumer-facing surface, with all the obligations that implies: design, documentation, versioning, security, rate limiting, and support. It is not a checkbox.
The failure mode I see most often is treating MCP as an export format. Someone points a generator at a 300-operation OpenAPI document and produces 300 tools. The agent now has an unusable tool list, picks the wrong one constantly, and the provider concludes that agents do not work. Tools are not endpoints. A good tool is a task an agent wants to accomplish, described in language a model can reason about, and there are usually far fewer of those than you have endpoints.
The other thing to plan for early is change. The tool name and description are load-bearing for the model in a way an endpoint path never was for a developer, and there is no version negotiation in the protocol to protect you. Renaming a tool is a breaking change even though nothing about your HTTP API moved.
