MCP's Python SDK refuses to start when two extensions claim the same method name

Two of the three guards look at one binding at a time. The core-method guard fires when `MethodBinding` is constructed, so a vendor package that tries to bind `tools/list` fails on that object ^[5]. An empty `protocol_versions` frozenset describes a method that cannot be used under any protocol version, and the SDK rejects it immediately ^[6]. An extension author can cover both of those inside their own package's tests ^[2].

The duplicate check cannot live there. In the sample, `CatalogSearch` and `ShadowSearch` each return the same binding, and neither class is invalid on its own; the `ValueError` arrives only when both are handed to one `MCPServer` ^[9][4]. That is an argument about where a test goes, not about how strict the SDK is. The post's author keeps ownership tests beside the composition root where optional packages are assembled ^[16].

The counterfactual is what makes ordering the issue. "MCP Python SDK extension method collisions are configuration defects, not runtime edge cases," the author wrote ^[14]. Under a last-write-wins registry, the handler that answers `com.example/catalog.search` would depend on extension order, and reordering configuration could silently change request behaviour without changing the client call ^[3]. Refusing to start takes order out of the question: swapping the two classes cannot turn the failure into success ^[10].

The positive path is the part I would copy. One valid configuration and three invalid ones sit in the same runnable sample, four in total ^[11][1], and the valid case goes through the SDK's in-memory client, so it opens no port and needs no external MCP host ^[11]. The handler returns `["mcp-0", "mcp-1"]` ^[12]. "Without that control case, a test could pass simply because every extension path was broken," the author wrote ^[13].

What the guards buy you depends on the list you pass to the constructor. The duplicate check runs during `MCPServer` construction, so it sees exactly the extensions given to that call and nothing discovered later ^[3]. Method names are public contract here because an extension adds behaviour to the same dispatch table the rest of the server uses ^[2], and the reverse-domain prefix is the mechanism that keeps a vendor name clear of core MCP names ^[8]. The sample pins `mcp==2.1.1` and targets protocol version `2026-07-28`, announced in the project's final MCP release post, so the checks are reproducible ^[7]. It reports nothing about the registries in any other MCP SDK, so a TypeScript or Go server's behaviour on the same collision would have to be read out of those docs.