diff --git a/TracyLua.hpp b/TracyLua.hpp
index 174498f1..5fa1a1be 100644
--- a/TracyLua.hpp
+++ b/TracyLua.hpp
@@ -121,6 +121,7 @@ namespace detail
 
 static inline int LuaZoneBegin( lua_State* L )
 {
+#ifndef TRACY_ON_DEMAND
     const uint32_t color = Color::DeepSkyBlue3;
 
     lua_Debug dbg;
@@ -163,11 +164,13 @@ static inline int LuaZoneBegin( lua_State* L )
     MemWrite( &item->zoneBegin.thread, GetThreadHandle() );
     MemWrite( &item->zoneBegin.srcloc, (uint64_t)ptr );
     tail.store( magic + 1, std::memory_order_release );
+#endif
     return 0;
 }
 
 static inline int LuaZoneBeginN( lua_State* L )
 {
+#ifndef TRACY_ON_DEMAND
     const uint32_t color = Color::DeepSkyBlue3;
 
     lua_Debug dbg;
@@ -214,11 +217,13 @@ static inline int LuaZoneBeginN( lua_State* L )
     MemWrite( &item->zoneBegin.thread, GetThreadHandle() );
     MemWrite( &item->zoneBegin.srcloc, (uint64_t)ptr );
     tail.store( magic + 1, std::memory_order_release );
+#endif
     return 0;
 }
 
 static inline int LuaZoneEnd( lua_State* L )
 {
+#ifndef TRACY_ON_DEMAND
     Magic magic;
     auto& token = s_token.ptr;
     auto& tail = token->get_tail_index();
@@ -233,11 +238,13 @@ static inline int LuaZoneEnd( lua_State* L )
 #endif
     MemWrite( &item->zoneEnd.thread, GetThreadHandle() );
     tail.store( magic + 1, std::memory_order_release );
+#endif
     return 0;
 }
 
 static inline int LuaZoneText( lua_State* L )
 {
+#ifndef TRACY_ON_DEMAND
     auto txt = lua_tostring( L, 1 );
     const auto size = strlen( txt );
 
@@ -252,11 +259,13 @@ static inline int LuaZoneText( lua_State* L )
     MemWrite( &item->zoneText.thread, GetThreadHandle() );
     MemWrite( &item->zoneText.text, (uint64_t)ptr );
     tail.store( magic + 1, std::memory_order_release );
+#endif
     return 0;
 }
 
 static inline int LuaZoneName( lua_State* L )
 {
+#ifndef TRACY_ON_DEMAND
     auto txt = lua_tostring( L, 1 );
     const auto size = strlen( txt );
 
@@ -271,11 +280,13 @@ static inline int LuaZoneName( lua_State* L )
     MemWrite( &item->zoneText.thread, GetThreadHandle() );
     MemWrite( &item->zoneText.text, (uint64_t)ptr );
     tail.store( magic + 1, std::memory_order_release );
+#endif
     return 0;
 }
 
 static inline int LuaMessage( lua_State* L )
 {
+#ifndef TRACY_ON_DEMAND
     auto txt = lua_tostring( L, 1 );
     const auto size = strlen( txt );
 
@@ -291,6 +302,7 @@ static inline int LuaMessage( lua_State* L )
     MemWrite( &item->message.thread, GetThreadHandle() );
     MemWrite( &item->message.text, (uint64_t)ptr );
     tail.store( magic + 1, std::memory_order_release );
+#endif
     return 0;
 }
 
diff --git a/TracyOpenGL.hpp b/TracyOpenGL.hpp
index 5248cfe3..80f00771 100644
--- a/TracyOpenGL.hpp
+++ b/TracyOpenGL.hpp
@@ -65,6 +65,7 @@ public:
         GLint bits;
         glGetQueryiv( GL_TIMESTAMP, GL_QUERY_COUNTER_BITS, &bits );
 
+#ifndef TRACY_ON_DEMAND
         const float period = 1.f;
         Magic magic;
         auto& token = s_token.ptr;
@@ -78,10 +79,12 @@ public:
         MemWrite( &item->gpuNewContext.context, m_context );
         MemWrite( &item->gpuNewContext.accuracyBits, (uint8_t)bits );
         tail.store( magic + 1, std::memory_order_release );
+#endif
     }
 
     void Collect()
     {
+#ifndef TRACY_ON_DEMAND
         ZoneScopedC( Color::Red4 );
 
         auto start = m_tail;
@@ -122,6 +125,7 @@ public:
             tail.store( magic + 1, std::memory_order_release );
             m_tail = ( m_tail + 1 ) % QueryCount;
         }
+#endif
     }
 
 private:
@@ -157,6 +161,7 @@ class GpuCtxScope
 public:
     tracy_force_inline GpuCtxScope( const SourceLocation* srcloc )
     {
+#ifndef TRACY_ON_DEMAND
         const auto queryId = s_gpuCtx.ptr->NextQueryId();
         glQueryCounter( s_gpuCtx.ptr->TranslateOpenGlQueryId( queryId ), GL_TIMESTAMP );
 
@@ -171,10 +176,12 @@ public:
         MemWrite( &item->gpuZoneBegin.queryId, uint16_t( queryId ) );
         MemWrite( &item->gpuZoneBegin.context, s_gpuCtx.ptr->GetId() );
         tail.store( magic + 1, std::memory_order_release );
+#endif
     }
 
     tracy_force_inline GpuCtxScope( const SourceLocation* srcloc, int depth )
     {
+#ifndef TRACY_ON_DEMAND
         const auto queryId = s_gpuCtx.ptr->NextQueryId();
         glQueryCounter( s_gpuCtx.ptr->TranslateOpenGlQueryId( queryId ), GL_TIMESTAMP );
 
@@ -193,10 +200,12 @@ public:
         tail.store( magic + 1, std::memory_order_release );
 
         s_profiler.SendCallstack( depth, thread );
+#endif
     }
 
     tracy_force_inline ~GpuCtxScope()
     {
+#ifndef TRACY_ON_DEMAND
         const auto queryId = s_gpuCtx.ptr->NextQueryId();
         glQueryCounter( s_gpuCtx.ptr->TranslateOpenGlQueryId( queryId ), GL_TIMESTAMP );
 
@@ -209,6 +218,7 @@ public:
         MemWrite( &item->gpuZoneEnd.queryId, uint16_t( queryId ) );
         MemWrite( &item->gpuZoneEnd.context, s_gpuCtx.ptr->GetId() );
         tail.store( magic + 1, std::memory_order_release );
+#endif
     }
 };
 
diff --git a/TracyVulkan.hpp b/TracyVulkan.hpp
index 161e5bc9..684eb886 100644
--- a/TracyVulkan.hpp
+++ b/TracyVulkan.hpp
@@ -97,6 +97,7 @@ public:
         vkQueueSubmit( queue, 1, &submitInfo, VK_NULL_HANDLE );
         vkQueueWaitIdle( queue );
 
+#ifndef TRACY_ON_DEMAND
         Magic magic;
         auto& token = s_token.ptr;
         auto& tail = token->get_tail_index();
@@ -109,6 +110,7 @@ public:
         MemWrite( &item->gpuNewContext.context, m_context );
         MemWrite( &item->gpuNewContext.accuracyBits, uint8_t( 0 ) );
         tail.store( magic + 1, std::memory_order_release );
+#endif
     }
 
     ~VkCtx()
@@ -118,6 +120,7 @@ public:
 
     void Collect( VkCommandBuffer cmdbuf )
     {
+#ifndef TRACY_ON_DEMAND
         ZoneScopedC( Color::Red4 );
 
         if( m_tail == m_head ) return;
@@ -158,6 +161,7 @@ public:
 
         m_tail += cnt;
         if( m_tail == QueryCount ) m_tail = 0;
+#endif
     }
 
 private:
@@ -192,6 +196,7 @@ public:
     tracy_force_inline VkCtxScope( const SourceLocation* srcloc, VkCommandBuffer cmdbuf )
         : m_cmdbuf( cmdbuf )
     {
+#ifndef TRACY_ON_DEMAND
         auto ctx = s_vkCtx.ptr;
         const auto queryId = ctx->NextQueryId();
         vkCmdWriteTimestamp( cmdbuf, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT, ctx->m_query, queryId );
@@ -207,11 +212,13 @@ public:
         MemWrite( &item->gpuZoneBegin.queryId, uint16_t( queryId ) );
         MemWrite( &item->gpuZoneBegin.context, ctx->GetId() );
         tail.store( magic + 1, std::memory_order_release );
+#endif
     }
 
     tracy_force_inline VkCtxScope( const SourceLocation* srcloc, VkCommandBuffer cmdbuf, int depth )
         : m_cmdbuf( cmdbuf )
     {
+#ifndef TRACY_ON_DEMAND
         const auto thread = GetThreadHandle();
 
         auto ctx = s_vkCtx.ptr;
@@ -231,10 +238,12 @@ public:
         tail.store( magic + 1, std::memory_order_release );
 
         s_profiler.SendCallstack( depth, thread );
+#endif
     }
 
     tracy_force_inline ~VkCtxScope()
     {
+#ifndef TRACY_ON_DEMAND
         auto ctx = s_vkCtx.ptr;
         const auto queryId = ctx->NextQueryId();
         vkCmdWriteTimestamp( m_cmdbuf, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT, ctx->m_query, queryId );
@@ -248,6 +257,7 @@ public:
         MemWrite( &item->gpuZoneEnd.queryId, uint16_t( queryId ) );
         MemWrite( &item->gpuZoneEnd.context, ctx->GetId() );
         tail.store( magic + 1, std::memory_order_release );
+#endif
     }
 
 private:
diff --git a/client/TracyLock.hpp b/client/TracyLock.hpp
index d3a7f3d1..da9f6142 100644
--- a/client/TracyLock.hpp
+++ b/client/TracyLock.hpp
@@ -22,6 +22,7 @@ public:
     {
         assert( m_id != std::numeric_limits<uint32_t>::max() );
 
+#ifndef TRACY_ON_DEMAND
         Magic magic;
         auto& token = s_token.ptr;
         auto& tail = token->get_tail_index();
@@ -31,6 +32,7 @@ public:
         MemWrite( &item->lockAnnounce.lckloc, (uint64_t)srcloc );
         MemWrite( &item->lockAnnounce.type, LockType::Lockable );
         tail.store( magic + 1, std::memory_order_release );
+#endif
     }
 
     Lockable( const Lockable& ) = delete;
@@ -38,6 +40,7 @@ public:
 
     tracy_force_inline void lock()
     {
+#ifndef TRACY_ON_DEMAND
         const auto thread = GetThreadHandle();
         {
             Magic magic;
@@ -51,9 +54,11 @@ public:
             MemWrite( &item->lockWait.type, LockType::Lockable );
             tail.store( magic + 1, std::memory_order_release );
         }
+#endif
 
         m_lockable.lock();
 
+#ifndef TRACY_ON_DEMAND
         {
             Magic magic;
             auto& token = s_token.ptr;
@@ -65,12 +70,14 @@ public:
             MemWrite( &item->lockObtain.time, Profiler::GetTime() );
             tail.store( magic + 1, std::memory_order_release );
         }
+#endif
     }
 
     tracy_force_inline void unlock()
     {
         m_lockable.unlock();
 
+#ifndef TRACY_ON_DEMAND
         Magic magic;
         auto& token = s_token.ptr;
         auto& tail = token->get_tail_index();
@@ -80,11 +87,13 @@ public:
         MemWrite( &item->lockRelease.thread, GetThreadHandle() );
         MemWrite( &item->lockRelease.time, Profiler::GetTime() );
         tail.store( magic + 1, std::memory_order_release );
+#endif
     }
 
     tracy_force_inline bool try_lock()
     {
         const auto ret = m_lockable.try_lock();
+#ifndef TRACY_ON_DEMAND
         if( ret )
         {
             Magic magic;
@@ -97,11 +106,13 @@ public:
             MemWrite( &item->lockObtain.time, Profiler::GetTime() );
             tail.store( magic + 1, std::memory_order_release );
         }
+#endif
         return ret;
     }
 
     tracy_force_inline void Mark( const SourceLocation* srcloc ) const
     {
+#ifndef TRACY_ON_DEMAND
         Magic magic;
         auto& token = s_token.ptr;
         auto& tail = token->get_tail_index();
@@ -111,6 +122,7 @@ public:
         MemWrite( &item->lockMark.thread, GetThreadHandle() );
         MemWrite( &item->lockMark.srcloc, (uint64_t)srcloc );
         tail.store( magic + 1, std::memory_order_release );
+#endif
     }
 
 private:
@@ -128,6 +140,7 @@ public:
     {
         assert( m_id != std::numeric_limits<uint32_t>::max() );
 
+#ifndef TRACY_ON_DEMAND
         Magic magic;
         auto& token = s_token.ptr;
         auto& tail = token->get_tail_index();
@@ -137,6 +150,7 @@ public:
         MemWrite( &item->lockAnnounce.lckloc, (uint64_t)srcloc );
         MemWrite( &item->lockAnnounce.type, LockType::SharedLockable );
         tail.store( magic + 1, std::memory_order_release );
+#endif
     }
 
     SharedLockable( const SharedLockable& ) = delete;
@@ -144,6 +158,7 @@ public:
 
     tracy_force_inline void lock()
     {
+#ifndef TRACY_ON_DEMAND
         const auto thread = GetThreadHandle();
         {
             Magic magic;
@@ -157,9 +172,11 @@ public:
             MemWrite( &item->lockWait.type, LockType::SharedLockable );
             tail.store( magic + 1, std::memory_order_release );
         }
+#endif
 
         m_lockable.lock();
 
+#ifndef TRACY_ON_DEMAND
         {
             Magic magic;
             auto& token = s_token.ptr;
@@ -171,12 +188,14 @@ public:
             MemWrite( &item->lockObtain.time, Profiler::GetTime() );
             tail.store( magic + 1, std::memory_order_release );
         }
+#endif
     }
 
     tracy_force_inline void unlock()
     {
         m_lockable.unlock();
 
+#ifndef TRACY_ON_DEMAND
         Magic magic;
         auto& token = s_token.ptr;
         auto& tail = token->get_tail_index();
@@ -186,11 +205,13 @@ public:
         MemWrite( &item->lockRelease.thread, GetThreadHandle() );
         MemWrite( &item->lockRelease.time, Profiler::GetTime() );
         tail.store( magic + 1, std::memory_order_release );
+#endif
     }
 
     tracy_force_inline bool try_lock()
     {
         const auto ret = m_lockable.try_lock();
+#ifndef TRACY_ON_DEMAND
         if( ret )
         {
             Magic magic;
@@ -203,11 +224,13 @@ public:
             MemWrite( &item->lockObtain.time, Profiler::GetTime() );
             tail.store( magic + 1, std::memory_order_release );
         }
+#endif
         return ret;
     }
 
     tracy_force_inline void lock_shared()
     {
+#ifndef TRACY_ON_DEMAND
         const auto thread = GetThreadHandle();
         {
             Magic magic;
@@ -221,9 +244,11 @@ public:
             MemWrite( &item->lockWait.type, LockType::SharedLockable );
             tail.store( magic + 1, std::memory_order_release );
         }
+#endif
 
         m_lockable.lock_shared();
 
+#ifndef TRACY_ON_DEMAND
         {
             Magic magic;
             auto& token = s_token.ptr;
@@ -235,12 +260,14 @@ public:
             MemWrite( &item->lockObtain.time, Profiler::GetTime() );
             tail.store( magic + 1, std::memory_order_release );
         }
+#endif
     }
 
     tracy_force_inline void unlock_shared()
     {
         m_lockable.unlock_shared();
 
+#ifndef TRACY_ON_DEMAND
         Magic magic;
         auto& token = s_token.ptr;
         auto& tail = token->get_tail_index();
@@ -250,11 +277,13 @@ public:
         MemWrite( &item->lockRelease.thread, GetThreadHandle() );
         MemWrite( &item->lockRelease.time, Profiler::GetTime() );
         tail.store( magic + 1, std::memory_order_release );
+#endif
     }
 
     tracy_force_inline bool try_lock_shared()
     {
         const auto ret = m_lockable.try_lock_shared();
+#ifndef TRACY_ON_DEMAND
         if( ret )
         {
             Magic magic;
@@ -267,11 +296,13 @@ public:
             MemWrite( &item->lockObtain.time, Profiler::GetTime() );
             tail.store( magic + 1, std::memory_order_release );
         }
+#endif
         return ret;
     }
 
     tracy_force_inline void Mark( const SourceLocation* srcloc ) const
     {
+#ifndef TRACY_ON_DEMAND
         Magic magic;
         auto& token = s_token.ptr;
         auto& tail = token->get_tail_index();
@@ -281,6 +312,7 @@ public:
         MemWrite( &item->lockMark.thread, GetThreadHandle() );
         MemWrite( &item->lockMark.srcloc, (uint64_t)srcloc );
         tail.store( magic + 1, std::memory_order_release );
+#endif
     }
 
 private:
diff --git a/client/TracyProfiler.hpp b/client/TracyProfiler.hpp
index a589ae72..7b1a102d 100644
--- a/client/TracyProfiler.hpp
+++ b/client/TracyProfiler.hpp
@@ -118,6 +118,7 @@ public:
 
     static tracy_force_inline void FrameMark()
     {
+#ifndef TRACY_ON_DEMAND
         Magic magic;
         auto& token = s_token.ptr;
         auto& tail = token->get_tail_index();
@@ -125,10 +126,12 @@ public:
         MemWrite( &item->hdr.type, QueueType::FrameMarkMsg );
         MemWrite( &item->frameMark.time, GetTime() );
         tail.store( magic + 1, std::memory_order_release );
+#endif
     }
 
     static tracy_force_inline void PlotData( const char* name, int64_t val )
     {
+#ifndef TRACY_ON_DEMAND
         Magic magic;
         auto& token = s_token.ptr;
         auto& tail = token->get_tail_index();
@@ -139,10 +142,12 @@ public:
         MemWrite( &item->plotData.type, PlotDataType::Int );
         MemWrite( &item->plotData.data.i, val );
         tail.store( magic + 1, std::memory_order_release );
+#endif
     }
 
     static tracy_force_inline void PlotData( const char* name, float val )
     {
+#ifndef TRACY_ON_DEMAND
         Magic magic;
         auto& token = s_token.ptr;
         auto& tail = token->get_tail_index();
@@ -153,10 +158,12 @@ public:
         MemWrite( &item->plotData.type, PlotDataType::Float );
         MemWrite( &item->plotData.data.f, val );
         tail.store( magic + 1, std::memory_order_release );
+#endif
     }
 
     static tracy_force_inline void PlotData( const char* name, double val )
     {
+#ifndef TRACY_ON_DEMAND
         Magic magic;
         auto& token = s_token.ptr;
         auto& tail = token->get_tail_index();
@@ -167,10 +174,12 @@ public:
         MemWrite( &item->plotData.type, PlotDataType::Double );
         MemWrite( &item->plotData.data.d, val );
         tail.store( magic + 1, std::memory_order_release );
+#endif
     }
 
     static tracy_force_inline void Message( const char* txt, size_t size )
     {
+#ifndef TRACY_ON_DEMAND
         Magic magic;
         auto& token = s_token.ptr;
         auto ptr = (char*)tracy_malloc( size+1 );
@@ -183,10 +192,12 @@ public:
         MemWrite( &item->message.thread, GetThreadHandle() );
         MemWrite( &item->message.text, (uint64_t)ptr );
         tail.store( magic + 1, std::memory_order_release );
+#endif
     }
 
     static tracy_force_inline void Message( const char* txt )
     {
+#ifndef TRACY_ON_DEMAND
         Magic magic;
         auto& token = s_token.ptr;
         auto& tail = token->get_tail_index();
@@ -196,29 +207,35 @@ public:
         MemWrite( &item->message.thread, GetThreadHandle() );
         MemWrite( &item->message.text, (uint64_t)txt );
         tail.store( magic + 1, std::memory_order_release );
+#endif
     }
 
     static tracy_force_inline void MemAlloc( const void* ptr, size_t size )
     {
+#ifndef TRACY_ON_DEMAND
         const auto thread = GetThreadHandle();
 
         s_profiler.m_serialLock.lock();
         SendMemAlloc( QueueType::MemAlloc, thread, ptr, size );
         s_profiler.m_serialLock.unlock();
+#endif
     }
 
     static tracy_force_inline void MemFree( const void* ptr )
     {
+#ifndef TRACY_ON_DEMAND
         const auto thread = GetThreadHandle();
 
         s_profiler.m_serialLock.lock();
         SendMemFree( QueueType::MemFree, thread, ptr );
         s_profiler.m_serialLock.unlock();
+#endif
     }
 
     static tracy_force_inline void MemAllocCallstack( const void* ptr, size_t size, int depth )
     {
 #ifdef TRACY_HAS_CALLSTACK
+#  ifndef TRACY_ON_DEMAND
         const auto thread = GetThreadHandle();
 
         rpmalloc_thread_initialize();
@@ -228,6 +245,7 @@ public:
         SendMemAlloc( QueueType::MemAllocCallstack, thread, ptr, size );
         SendCallstackMemory( callstack );
         s_profiler.m_serialLock.unlock();
+#  endif
 #else
         MemAlloc( ptr, size );
 #endif
@@ -236,6 +254,7 @@ public:
     static tracy_force_inline void MemFreeCallstack( const void* ptr, int depth )
     {
 #ifdef TRACY_HAS_CALLSTACK
+#  ifndef TRACY_ON_DEMAND
         const auto thread = GetThreadHandle();
 
         rpmalloc_thread_initialize();
@@ -245,6 +264,7 @@ public:
         SendMemFree( QueueType::MemFreeCallstack, thread, ptr );
         SendCallstackMemory( callstack );
         s_profiler.m_serialLock.unlock();
+#  endif
 #else
         MemFree( ptr );
 #endif
diff --git a/client/TracyScoped.hpp b/client/TracyScoped.hpp
index 21e5fcd2..b81be485 100644
--- a/client/TracyScoped.hpp
+++ b/client/TracyScoped.hpp
@@ -17,6 +17,7 @@ class ScopedZone
 public:
     tracy_force_inline ScopedZone( const SourceLocation* srcloc )
     {
+#ifndef TRACY_ON_DEMAND
         const auto thread = GetThreadHandle();
         m_thread = thread;
         Magic magic;
@@ -34,10 +35,12 @@ public:
         MemWrite( &item->zoneBegin.thread, thread );
         MemWrite( &item->zoneBegin.srcloc, (uint64_t)srcloc );
         tail.store( magic + 1, std::memory_order_release );
+#endif
     }
 
     tracy_force_inline ScopedZone( const SourceLocation* srcloc, int depth )
     {
+#ifndef TRACY_ON_DEMAND
         const auto thread = GetThreadHandle();
         m_thread = thread;
         Magic magic;
@@ -57,10 +60,12 @@ public:
         tail.store( magic + 1, std::memory_order_release );
 
         s_profiler.SendCallstack( depth, thread );
+#endif
     }
 
     tracy_force_inline ~ScopedZone()
     {
+#ifndef TRACY_ON_DEMAND
         Magic magic;
         auto& token = s_token.ptr;
         auto& tail = token->get_tail_index();
@@ -75,10 +80,12 @@ public:
 #endif
         MemWrite( &item->zoneEnd.thread, m_thread );
         tail.store( magic + 1, std::memory_order_release );
+#endif
     }
 
     tracy_force_inline void Text( const char* txt, size_t size )
     {
+#ifndef TRACY_ON_DEMAND
         Magic magic;
         auto& token = s_token.ptr;
         auto ptr = (char*)tracy_malloc( size+1 );
@@ -90,10 +97,12 @@ public:
         MemWrite( &item->zoneText.thread, m_thread );
         MemWrite( &item->zoneText.text, (uint64_t)ptr );
         tail.store( magic + 1, std::memory_order_release );
+#endif
     }
 
     tracy_force_inline void Name( const char* txt, size_t size )
     {
+#ifndef TRACY_ON_DEMAND
         Magic magic;
         auto& token = s_token.ptr;
         auto ptr = (char*)tracy_malloc( size+1 );
@@ -105,6 +114,7 @@ public:
         MemWrite( &item->zoneText.thread, m_thread );
         MemWrite( &item->zoneText.text, (uint64_t)ptr );
         tail.store( magic + 1, std::memory_order_release );
+#endif
     }
 
 private: