Anthropic::Middleware

HTTP middleware system for inspecting, rewriting, or short-circuiting requests and responses around the SDK's terminal HTTP send.

A middleware is any type that includes Anthropic::Middleware and implements #call(request, nxt). The chain runs once per HTTP attempt, inside the SDK's retry loop, so a middleware observes every retry and may call nxt multiple times to implement custom retry logic.

nxt.call(request) returns an APIResponse for every HTTP status (4xx/5xx do not raise inside the chain — typed error raising happens after the chain returns). Connection-level failures (APITimeoutError, APIConnectionError) do raise from nxt.call.

Limitation: only buffered JSON requests (GET/POST/DELETE) run through the chain. Streaming (messages.stream / open_stream / SSE endpoints), raw downloads, and multipart uploads bypass middleware entirely.

Registration: Anthropic::Client.new(middleware: [my_mw]) (outermost first).

class LoggingMiddleware
  include Anthropic::Middleware

  def call(request : Anthropic::APIRequest, nxt : Anthropic::MiddlewareNext) : Anthropic::APIResponse
    puts "-> #{request.method} #{request.path}"
    response = nxt.call(request)
    puts "<- #{response.status}"
    response
  end
end

client = Anthropic::Client.new(middleware: [LoggingMiddleware.new])