A developer reports that an agent using an MCP server to query an internal ticketing system is failing with 'connection refused' errors on every invocation, despite the MCP server process running and responding to manual HTTP requests on the expected port. The agent's MCP configuration file specifies the correct host and port. What is the most likely cause?

When an MCP server is reachable by manual HTTP request but the agent receives 'connection refused' errors, the distinction is likely in authentication. Manual requests from a developer's terminal often include pre-configured credentials (environment variables, stored tokens, or explicit header flags), while the agent's MCP client uses only what is specified in its configuration file. A missing or incorrect authentication credential causes the server to reject the connection before any tool invocation occurs.

The diagnostic step is to compare the headers in a successful manual request against the headers the agent's MCP client sends. The most common fix is adding the required Authorization or X-API-Key header to the MCP server entry in the agent's configuration.

Option A (server not supporting the tool) would manifest as a different error: a successful connection followed by a tool-not-found or method-not-allowed response, not 'connection refused' before a connection is established.

Option C (tool not in allow list) is an agent-side enforcement issue: the agent's own allow list would prevent it from calling the tool, but the error message would come from the agent runtime, not the MCP server, and would not appear as 'connection refused.'

Option D (schema incompatibility) would result in errors at tool negotiation time after the connection is established, not at connection establishment.

Auth misconfiguration is the primary cause of MCP server 'connection refused' errors when the server itself is running correctly.