The Claude connector gates that fail before your code runs: DNS, IPv4 and one Location header

Three lookups and one curl. That is the entire first gate ^[15], and none of it needs a deploy, which is the argument for running it before the first tool is written. The checklist's own framing is that desk tests prove almost nothing, because Claude dials in from Anthropic's servers ^[2]; the answer that decides whether you ship is the one a resolver outside your network returns ^[7].

The expensive fault is the redirect. RFC 9110 section 15.4 strips the Authorization header when a 301, 302, 307 or 308 crosses to a different host, so the 401 the user sees is generated by the redirect target rather than by your MCP endpoint ^[5]. Claude surfaces that as an auth failure, which sends the developer straight to the identity gate while the defect sits in the network gate ^[16]. The same checklist also lists cross-host redirects among the reasons your access log is empty while Claude reports that it could not reach the server ^[6], so both of the signals available to you point somewhere other than the Location header.

The identity gate has the same shape: a config detail that fails quietly. Claude selects a client ID metadata document only when the metadata advertises both `client_id_metadata_document_supported: true` and `token_endpoint_auth_methods_supported: ["none"]`. Miss the second and it falls back to dynamic registration and reports an incompatible auth server, with a perfectly good CIMD sitting right there ^[10]. `registration_endpoint` has to be absent rather than `null`, because `null` fails schema validation instead of being ignored ^[11].

Gate three is a human reviewer, and it has an automatic fail that plenty of servers ship on day one: a single tool whose schema accepts GET, POST and DELETE in one enum. Read and write must be separate tools, and no amount of description text rescues the combined version ^[12]. Worth knowing before anyone plans a launch date: the submission portal lives in Claude.ai org settings and requires a Team or Enterprise plan, so individual accounts cannot submit at all ^[14].

Gate four reverses advice many teams have already acted on. Tool search is on by default in Claude Code, loading only tool names and server instructions at session start and fetching full schemas on demand, and Anthropic's reference says additional servers have minimal impact on the window, with no fixed per-server tool cap ^[1]. If you split one server into three to protect context, that was work done for nothing.