decision frameworks

Compare & Decide

Every piece ends with a verdict and the conditions under which it flips. If a comparison can't tell you when to choose the loser, it isn't done.

FastMCP 2.x vs the official MCP Python SDK: which should you use?

Untangling the FastMCP naming confusion, what the official mcp package gives you, what FastMCP 2.x adds, and a clear verdict for each kind of project.

verdict Start with the official mcp SDK — FastMCP 1.0 is already inside it and it tracks the spec first. Reach for FastMCP 2.x when you concretely need its extras: auth helpers, proxying, server composition, or its testing utilities.

MCP vs OpenAPI: why not just give the model your API spec?

Auto-generating MCP tools from an OpenAPI spec vs hand-designing task-shaped tools: context cost, abstraction leaks, and when each approach actually works.

verdict Auto-generate from OpenAPI only for small, internal specs you fully control; for anything user-facing or large, hand-design a few task-shaped tools — endpoints are for programs, tools are for models.

MCP transports: stdio vs Streamable HTTP (and what happened to SSE)

How to choose between stdio and Streamable HTTP for your MCP server, why SSE is deprecated, and how the 2026-07-28 stateless core changes the design maths.

verdict Local and personal tools: stdio. Anything remote, shared, or ChatGPT-facing: Streamable HTTP. SSE is deprecated — migrate. And design stateless now, because the 2026-07-28 spec makes statelessness the core model.

MCP vs function calling: what's actually different?

Function calling is the model-side mechanism; MCP standardises how tools are packaged, discovered, and shared across apps. How they compose — and when each wins.

verdict They compose, not compete: function calling is how the model asks for a tool, MCP is how tools reach the model. Use plain function calling for one app with a few tools; use MCP the moment tools need to be reused across clients or teams.

Build your own MCP server or use an existing one?

How to decide between installing a third-party MCP server and writing your own: a vetting checklist for existing servers, and the cases where building is faster.

verdict Reuse a well-maintained server for commodity integrations after a real vetting pass; build your own for anything touching your systems, your data, or your policies — it is usually faster than it looks and safer than it sounds.

MCP vs A2A: the agent protocol stack, sorted

MCP connects an agent to its tools and context; A2A connects agents to each other. Where AG-UI and AGENTS.md fit, one diagram, and how to pick by boundary.

verdict Not rivals but layers: MCP for the agent-to-tools boundary, A2A for the agent-to-agent boundary. Name the boundary you are crossing and the protocol picks itself.