diff --git a/lib/system/gc.nim b/lib/system/gc.nim
index fe4b409035..e0db3fba43 100644
--- a/lib/system/gc.nim
+++ b/lib/system/gc.nim
@@ -292,6 +292,20 @@ proc initGC() =
     when useMarkForDebug or useBackupGc:
       init(gch.marked)
 
+var
+  localGcInitialized {.rtlThreadVar.}: bool
+
+proc setupForeignThreadGc*() =
+  ## call this if you registered a callback that will be run from a thread not
+  ## under your control. This has a cheap thread-local guard, so the GC for
+  ## this thread will only be initialized once per thread, no matter how often
+  ## it is called.
+  if not localGcInitialized:
+    localGcInitialized = true
+    var stackTop {.volatile.}: pointer
+    setStackBottom(addr(stackTop))
+    initGC()
+
 when useMarkForDebug or useBackupGc:
   type
     TGlobalMarkerProc = proc () {.nimcall, benign.}
diff --git a/lib/system/gc_ms.nim b/lib/system/gc_ms.nim
index 242ca16084..0eafd3255d 100644
--- a/lib/system/gc_ms.nim
+++ b/lib/system/gc_ms.nim
@@ -168,6 +168,20 @@ proc initGC() =
       Init(gch.allocated)
       init(gch.marked)
 
+var
+  localGcInitialized {.rtlThreadVar.}: bool
+
+proc setupForeignThreadGc*() =
+  ## call this if you registered a callback that will be run from a thread not
+  ## under your control. This has a cheap thread-local guard, so the GC for
+  ## this thread will only be initialized once per thread, no matter how often
+  ## it is called.
+  if not localGcInitialized:
+    localGcInitialized = true
+    var stackTop {.volatile.}: pointer
+    setStackBottom(addr(stackTop))
+    initGC()
+
 proc forAllSlotsAux(dest: pointer, n: ptr TNimNode, op: TWalkOp) {.benign.} =
   var d = cast[ByteAddress](dest)
   case n.kind