diff --git a/CHANGELOG.md b/CHANGELOG.md
index 1b2ca20a..40a1315c 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -1,5 +1,9 @@
 _You can download previous versions [here](https://github.com/AliSoftware/OHHTTPStubs/tags) and latest version [here](https://github.com/AliSoftware/OHHTTPStubs/zipball/master) (ZIP files generated by GitHub on the fly)_
 
+## [4.0.1](https://github.com/AliSoftware/OHHTTPStubs/releases/tag/4.0.1)
+
+* Fix threading in `NSURLProtocol` subclass calling `NSURLProtocolClient` callbacks from wrong thread. ([@nsprogrammer](https://github.com/nsprogrammer), [#96](https://github.com/AliSoftware/OHHTTPStubs/pull/96))
+
 ## [4.0.0](https://github.com/AliSoftware/OHHTTPStubs/releases/tag/4.0.0) — Improvements for Swift 
 
 * Annotated the library with _nullability_ attributes to generate a better API when used in Swift
diff --git a/OHHTTPStubs/Sources/OHHTTPStubs.m b/OHHTTPStubs/Sources/OHHTTPStubs.m
index c18e67c2..4d36ae69 100644
--- a/OHHTTPStubs/Sources/OHHTTPStubs.m
+++ b/OHHTTPStubs/Sources/OHHTTPStubs.m
@@ -269,6 +269,8 @@ - (OHHTTPStubsDescriptor*)firstStubPassingTestForRequest:(NSURLRequest*)request
 @interface OHHTTPStubsProtocol()
 @property(assign) BOOL stopped;
 @property(strong) OHHTTPStubsDescriptor* stub;
+@property(assign) CFRunLoopRef clientRunLoop;
+- (void)executeOnClientRunLoopAfterDelay:(NSTimeInterval)delayInSeconds block:(dispatch_block_t)block;
 @end
 
 @implementation OHHTTPStubsProtocol
@@ -298,6 +300,7 @@ - (NSCachedURLResponse *)cachedResponse
 
 - (void)startLoading
 {
+    self.clientRunLoop = CFRunLoopGetCurrent();
     NSURLRequest* request = self.request;
     id<NSURLProtocolClient> client = self.client;
     
@@ -354,16 +357,16 @@ - (void)startLoading
         if (((responseStub.statusCode > 300) && (responseStub.statusCode < 400)) && redirectLocationURL)
         {
             NSURLRequest* redirectRequest = [NSURLRequest requestWithURL:redirectLocationURL];
-            execute_after(responseStub.requestTime, ^{
+            [self executeOnClientRunLoopAfterDelay:responseStub.requestTime block:^{
                 if (!self.stopped)
                 {
                     [client URLProtocol:self wasRedirectedToRequest:redirectRequest redirectResponse:urlResponse];
                 }
-            });
+            }];
         }
         else
         {
-            execute_after(responseStub.requestTime,^{
+            [self executeOnClientRunLoopAfterDelay:responseStub.requestTime block:^{
                 if (!self.stopped)
                 {
                     [client URLProtocol:self didReceiveResponse:urlResponse cacheStoragePolicy:NSURLCacheStorageNotAllowed];
@@ -386,16 +389,16 @@ - (void)startLoading
                          }
                      }];
                 }
-            });
+            }];
         }
     } else {
         // Send the canned error
-        execute_after(responseStub.responseTime, ^{
+        [self executeOnClientRunLoopAfterDelay:responseStub.responseTime block:^{
             if (!self.stopped)
             {
                 [client URLProtocol:self didFailWithError:responseStub.error];
             }
-        });
+        }];
     }
 }
 
@@ -471,10 +474,10 @@ - (void) streamDataForClient:(id<NSURLProtocolClient>)client
         if (chunkSizeToRead == 0)
         {
             // Nothing to read at this pass, but probably later
-            execute_after(timingInfo.slotTime, ^{
+            [self executeOnClientRunLoopAfterDelay:timingInfo.slotTime block:^{
                 [self streamDataForClient:client fromStream:inputStream
                                timingInfo:timingInfo completion:completion];
-            });
+            }];
         } else {
             uint8_t* buffer = (uint8_t*)malloc(sizeof(uint8_t)*chunkSizeToRead);
             NSInteger bytesRead = [inputStream read:buffer maxLength:chunkSizeToRead];
@@ -483,11 +486,11 @@ - (void) streamDataForClient:(id<NSURLProtocolClient>)client
                 NSData * data = [NSData dataWithBytes:buffer length:bytesRead];
                 // Wait for 'slotTime' seconds before sending the chunk.
                 // If bytesRead < chunkSizePerSlot (because we are near the EOF), adjust slotTime proportionally to the bytes remaining
-                execute_after(((double)bytesRead / (double)chunkSizeToRead) * timingInfo.slotTime, ^{
+                [self executeOnClientRunLoopAfterDelay:((double)bytesRead / (double)chunkSizeToRead) * timingInfo.slotTime block:^{
                     [client URLProtocol:self didLoadData:data];
                     [self streamDataForClient:client fromStream:inputStream
                                    timingInfo:timingInfo completion:completion];
-                });
+                }];
             }
             else
             {
@@ -515,10 +518,13 @@ - (void) streamDataForClient:(id<NSURLProtocolClient>)client
 // Delayed execution utility methods
 /////////////////////////////////////////////
 
-static void execute_after(NSTimeInterval delayInSeconds, dispatch_block_t block)
+- (void)executeOnClientRunLoopAfterDelay:(NSTimeInterval)delayInSeconds block:(dispatch_block_t)block
 {
     dispatch_time_t popTime = dispatch_time(DISPATCH_TIME_NOW, (int64_t)(delayInSeconds * NSEC_PER_SEC));
-    dispatch_after(popTime, dispatch_get_global_queue(DISPATCH_QUEUE_PRIORITY_DEFAULT, 0), block);
+    dispatch_after(popTime, dispatch_get_global_queue(DISPATCH_QUEUE_PRIORITY_DEFAULT, 0), ^{
+        CFRunLoopPerformBlock(self.clientRunLoop, kCFRunLoopDefaultMode, block);
+        CFRunLoopWakeUp(self.clientRunLoop);
+    });
 }
 
 @end