claude.ai, Claude Desktop, mobile, and Cowork connectors

Support

Protocol revision

Not published. Claude still runs the initialize handshake, and Anthropic says support for 2026-07-28 is rolling out across Claude products.

Caveats and sources

Each one names the column it qualifies and cites the page it comes from, which is often not the page in the row's primary source.

stdio

Local servers reach Claude Desktop as MCPB desktop extensions, which Anthropic documents for Team and Enterprise plans. The other three surfaces have no local-server path. Source →

Legacy HTTP+SSE

Anthropic documents the hosted connector runtime as taking the URL of a remote MCP server without naming a transport. The Messages API connector and Managed Agents both document SSE, but those are different runtimes, so their support doesn't transfer. Source →

Metadata document

Claude selects a metadata document only when your authorization server metadata advertises both client_id_metadata_document_supported and none in token_endpoint_auth_methods_supported. Anthropic recommends it over dynamic registration for busy servers, because registration mints a new client on every fresh connection. Source →

Pre-registered client ID

Custom connectors only, where an admin supplies OAuth client credentials to scope the client to one organization. Directory connectors use a single shared OAuth application per connector, with no per-organization client. Source →

Custom headers

Request header authentication is in beta and rolling out by request — Anthropic's own wording is to contact them for early access, so don't build against it as generally available. Header names come from an allowlist of standard authentication and routing names, you can add up to four, an organization administrator enters them, and Claude sends each value exactly as typed without adding a scheme. Authorization can't be set on a connection that also uses OAuth. Source →

Redirects

If your registered URL redirects to a different host, the Authorization header is dropped on the redirect. Anthropic names this as the reason a server can work in Claude Code or MCP Inspector and fail on claude.ai: a local client fails fast on the redirect, while the hosted runtime follows it and loses the credential. Source →

Client credentials grant

A machine-to-machine client_credentials grant with no user in the loop isn't supported, because every connection requires user consent. Enterprise Managed Auth is the consent-free path, it's in beta on Team and Enterprise plans, and it excludes dynamic client registration. Source →