What is MCP, and when would you build an MCP server instead of just wrapping a REST API as a tool?
Answer it out loud first, then compare. That's what makes it stick.
MCP is an open protocol that lets any compatible AI client discover and call tools, read resources and use prompts from a server, using a standard JSON-RPC format. I build a server when the same capability should work across several clients or agents, or the customer wants one governed place for auth and audit. For a single app with two internal functions, a plain tool definition is simpler. MCP buys reuse and standardization, not magic.
- MCP = standard interface: tools, resources, prompts
- Server exposes, client or host connects
- Build a server for reuse across clients and central governance
- Plain function tools for one app and few tools
- Security still lives in the host and server
What it is. The Model Context Protocol standardizes how AI applications connect to external systems. A server exposes capabilities; a client inside a host application connects to it and passes what it finds to the model. Tools are model-controlled functions with a name, description and JSON input schema (and optionally an output schema); servers can also expose resources (data) and prompts (templates).
The question is about trade-offs, so give both sides.
Build an MCP server when
- Several clients need the same capability (an internal assistant, an IDE, a support agent), so you write the integration once.
- The customer wants a governed choke point: one place for authentication, rate limits, logging and an allowlist, owned by their platform team.
- The system is large and you want tool definitions versioned with the service that owns the data.
- You are shipping to a customer who already uses MCP-capable clients.
Use plain function tools when
- One application, a handful of tools, and no reuse.
- You need tight control over latency and the extra network hop is not worth it.
- Security review will be faster for an in-process function than a new network service.
Cautions that show seniority
- MCP does not remove the need for good tool design; a server exposing 60 raw endpoints is still a bad toolset.
- Tool annotations (readOnly, destructive, idempotent hints) are only hints; clients should not trust them from untrusted servers.
- Every tool definition from every connected server is input tokens, so tool count matters (see the tool search card).
- The spec says there should always be a human able to deny tool invocations, so the approval UI belongs in the host.
Anchor. Anthropic's FDE postings list shipping MCP servers, sub-agents and skills inside customer environments, so expect a hands-on follow-up.
- How does a client discover which tools a server has?
- Who is responsible for approving a destructive tool call in MCP?
- What changes if the server runs locally over stdio versus remotely over HTTP?
- Forward Deployed Engineer Interview Questions: 50 Proven Answers (prep, 2026-10-07) Lists 'What is MCP, and when would you build an MCP server rather than a REST wrapper?' and ties MCP servers to the Anthropic FDE job description.
- Tools - Model Context Protocol specification (docs, 2026-07-28) Tools are model-controlled with input and output schemas; humans should be able to deny invocations; annotations are untrusted unless from trusted servers.