diff --git a/ffi/event_thread.nim b/ffi/event_thread.nim
index 21cc6c2..21ccdb9 100644
--- a/ffi/event_thread.nim
+++ b/ffi/event_thread.nim
@@ -116,20 +116,24 @@ proc eventRun[T](ctx: ptr FFIContext[T]) {.async.} =
   var hb = HeartbeatMonitor.init(ctx)
   var notifiedStuck = false # latched forever — eventQueueStuck is sticky terminal.
 
-  while ctx.running.load():
+  # Keep draining after `running` flips false until the FFI thread has exited, so
+  # events emitted by an async {.ffiDtor.} teardown are still dispatched.
+  while ctx.running.load() or not ctx.ffiThreadExited.load():
     # Wake on enqueue or tick — whichever first.
     discard await ctx.eventQueueSignal.wait().withTimeout(EventThreadTickInterval)
 
     ctx.drainEventQueue()
 
-    # Fire after drain so reg.lock is free — FFI-thread would deadlock here.
-    if not notifiedStuck and ctx.eventQueueStuck.load():
-      onNotResponding(ctx)
-      notifiedStuck = true
+    # Liveness only applies while running; skip it during the teardown drain.
+    if ctx.running.load():
+      # Fire after drain so reg.lock is free — FFI-thread would deadlock here.
+      if not notifiedStuck and ctx.eventQueueStuck.load():
+        onNotResponding(ctx)
+        notifiedStuck = true
+      hb.check(ctx)
 
-    if not ctx.running.load():
-      break
-    hb.check(ctx)
+  # Catch anything enqueued between the last drain and the FFI thread's exit.
+  ctx.drainEventQueue()
 
 proc eventThreadBody[T](ctx: ptr FFIContext[T]) {.thread.} =
   ## Drains the event queue and runs the FFI-thread heartbeat check.
diff --git a/ffi/ffi_context.nim b/ffi/ffi_context.nim
index 5f3407d..751e63c 100644
--- a/ffi/ffi_context.nim
+++ b/ffi/ffi_context.nim
@@ -37,6 +37,9 @@ type FFIContext*[T] = object
   ffiHeartbeat*: Atomic[int64]
     # advanced each FFI-thread loop; event thread reads for liveness
   eventQueueStuck*: Atomic[bool] # sticky overflow flag
+  ffiThreadExited*: Atomic[bool]
+    # set once the FFI thread (including any async {.ffiDtor.} teardown) is done;
+    # keeps the event thread draining until then so teardown-emitted events land
   running: Atomic[bool] # To control when the threads are running
   registeredRequests: ptr Table[cstring, FFIRequestProc]
   requestTimeouts: ptr Table[cstring, int]
@@ -58,6 +61,19 @@ const
   DefaultRequestTimeout* = 5.seconds
     # Finite fallback (issue #93) so a wedged handler can't hang a caller forever.
 
+type FFITeardownProc*[T] = proc(lib: ptr T): Future[void] {.async.}
+
+proc ffiTeardownHook*[T](): var FFITeardownProc[T] =
+  ## Per-library teardown slot (one `{.global.}` per `T`), assigned at module init
+  ## by a non-empty `{.ffiDtor.}` and awaited by the FFI thread before it exits.
+  ##
+  ## A runtime slot, not an overloaded `ffiTeardown` resolved via `mixin`: the
+  ## constructor force-instantiates the FFI thread body before the dtor (declared
+  ## later in the source) is visible, so an overload would bind the no-op default
+  ## and the teardown would silently never run.
+  var hook {.global.}: FFITeardownProc[T]
+  hook
+
 include ./event_thread
 include ./ffi_thread
 
@@ -113,6 +129,7 @@ proc initContextResources*[T](ctx: ptr FFIContext[T]): Result[void, string] =
   initEventQueue(ctx[].eventQueue)
   ctx.ffiHeartbeat.store(0)
   ctx.eventQueueStuck.store(false)
+  ctx.ffiThreadExited.store(false)
   ctx.defaultRequestTimeout = DefaultRequestTimeout
 
   var success = false
@@ -180,9 +197,14 @@ proc signalStop*[T](ctx: ptr FFIContext[T]): Result[void, string] =
     error "failed to signal eventQueueSignal in signalStop", error = error
   ok()
 
-## Bound on how long clearContext waits for the FFI thread to exit before
-## leaking ctx rather than hanging the caller.
-const ThreadExitTimeout* = 1500.milliseconds
+## Bound on how long stopAndJoinThreads waits for a worker thread to exit before
+## leaking ctx rather than hanging the caller. Configurable because an async
+## `{.ffiDtor.}` runs its teardown (e.g. `switch.stop()` over many live
+## connections) on the FFI thread before it exits — a graceful shutdown can
+## outlast the default, and being cut short leaks the context instead of waiting.
+## Override at compile time with `-d:ffiThreadExitTimeoutMs=<ms>`.
+const ThreadExitTimeoutMs* {.intdefine: "ffiThreadExitTimeoutMs".} = 1500
+const ThreadExitTimeout* = ThreadExitTimeoutMs.milliseconds
 
 proc stopAndJoinThreads*[T](ctx: ptr FFIContext[T]): Result[void, string] =
   ## On timeout, returns err and skips remaining joins (leaves threads live).
diff --git a/ffi/ffi_thread.nim b/ffi/ffi_thread.nim
index 9f6eac2..f98834b 100644
--- a/ffi/ffi_thread.nim
+++ b/ffi/ffi_thread.nim
@@ -150,6 +150,12 @@ proc ffiThreadBody[T](ctx: ptr FFIContext[T]) {.thread.} =
     onFFIThread = false
     # Free handle refs on the FFI thread that allocated them (refc heap is thread-local).
     ctx[].handles.releaseAll()
+    # Teardown has run and no more events will be emitted from this thread; let
+    # the event thread stop draining and exit. Wake it so it notices without
+    # waiting a full tick.
+    ctx.ffiThreadExited.store(true)
+    ctx.eventQueueSignal.fireSync().isOkOr:
+      error "failed to wake event thread on FFI thread exit", err = error
     # Unblocks destroyFFIContext's bounded wait so cleanup can proceed.
     let fireRes = ctx.threadExitSignal.fireSync()
     if fireRes.isErr():
@@ -210,4 +216,15 @@ proc ffiThreadBody[T](ctx: ptr FFIContext[T]) {.thread.} =
       except CatchableError as e:
         error "draining pending FFI requests on shutdown raised", error = e.msg
 
+    # In-flight requests drained; run the library's async shutdown (e.g.
+    # `switch.stop()`) on this event loop before the thread joins. Only if a
+    # `{.ffiDtor.}` registered a hook and a request populated `myLib`. Exceptions
+    # are logged, never propagated: the thread must still fire threadExitSignal.
+    let teardown = ffiTeardownHook[T]()
+    if not teardown.isNil() and not ctx.myLib.isNil():
+      try:
+        await teardown(ctx.myLib)
+      except CatchableError as e:
+        error "library teardown raised on shutdown", error = e.msg
+
   waitFor ffiRun(ctx)
diff --git a/ffi/internal/ffi_macro.nim b/ffi/internal/ffi_macro.nim
index 69be972..6b263b9 100644
--- a/ffi/internal/ffi_macro.nim
+++ b/ffi/internal/ffi_macro.nim
@@ -1575,26 +1575,38 @@ macro ffiCtor*(args: varargs[untyped]): untyped =
   return stmts
 
 macro ffiDtor*(args: varargs[untyped]): untyped =
-  ## Defines a C-exported destructor that tears down the FFIContext after the
-  ## body runs.
+  ## Defines a C-exported destructor that tears down the FFIContext.
   ##
-  ## The annotated proc must have exactly one parameter of the library type.
-  ## The body contains any library-level cleanup to run before context teardown.
+  ## The annotated proc must have exactly one parameter of the library type. It
+  ## may be sync (no return type) or async (`Future[void]`) — an async dtor can
+  ## `await` a graceful library shutdown (e.g. `switch.stop()`) whose futures
+  ## live on the FFI event loop.
   ##
   ## The wire format follows the library default and can be overridden with
   ## `{.ffiDtor: "abi = c".}` / `{.ffiDtor: "abi = cbor".}`.
   ##
-  ## Example:
-  ##   proc waku_destroy*(w: Waku) {.ffiDtor.} =
-  ##     w.cleanup()
+  ## Example (sync):
+  ##   proc echo_destroy*(e: Echo) {.ffiDtor.} =
+  ##     e.close()
+  ##
+  ## Example (async):
+  ##   proc waku_destroy*(w: Waku): Future[void] {.ffiDtor.} =
+  ##     await w.stop()
   ##
   ## The generated C-exported proc has the signature:
   ##   int waku_destroy(void* ctx)
   ##
-  ## It extracts the library value from ctx, runs the body, then calls
-  ## destroyFFIContext to tear down the FFI thread and free the context.
+  ## A non-empty body is lifted into an async impl registered in the library's
+  ## `ffiTeardownHook` slot; the FFI thread awaits it on its own event loop, after
+  ## draining in-flight requests and just before it exits — so the body runs on
+  ## the worker thread, not the host (calling) thread. The C wrapper signals the
+  ## thread to stop and blocks (up to `ThreadExitTimeout`) until it, and the
+  ## teardown, finish, then frees the context. An empty/`discard` body registers
+  ## no hook.
+  ##
   ## Returns RET_OK on success, RET_ERR on failure (null/invalid ctx, or
-  ## destroyFFIContext failure).
+  ## destroyFFIContext failure — e.g. a teardown that outlasts ThreadExitTimeout,
+  ## which leaks the context rather than hanging the caller).
 
   requireBeforeGenBindings("`.ffiDtor.`")
   requireLibraryDeclared("`.ffiDtor.`")
@@ -1612,6 +1624,18 @@ macro ffiDtor*(args: varargs[untyped]): untyped =
   let libParamName = formalParams[1][0]
   let libTypeName = formalParams[1][1]
 
+  # A dtor is sync (no return type) or async (`Future[void]`); reject anything
+  # else up front rather than emitting an obscure downstream error.
+  let retTypeNode = formalParams[0]
+  let retIsFutureVoid =
+    retTypeNode.kind == nnkBracketExpr and $retTypeNode[0] == "Future" and
+    retTypeNode.len == 2 and $retTypeNode[1] == "void"
+  if retTypeNode.kind != nnkEmpty and not retIsFutureVoid:
+    error(
+      "ffiDtor: proc must return nothing (sync) or Future[void] (async), got: " &
+        retTypeNode.repr
+    )
+
   let procNameStr = block:
     let raw = $procName
     if raw.endsWith("*"):
@@ -1639,16 +1663,26 @@ macro ffiDtor*(args: varargs[untyped]): untyped =
     if ctx.isNil or cast[ptr FFIContext[`libTypeName`]](ctx)[].myLib.isNil:
       return RET_ERR
 
-  ffiBody.add quote do:
-    let `libParamName` = cast[ptr FFIContext[`libTypeName`]](ctx)[].myLib[]
-
   let isNoop =
     bodyNode.kind == nnkEmpty or (
       bodyNode.kind == nnkStmtList and bodyNode.len == 1 and
       bodyNode[0].kind == nnkDiscardStmt
     )
-  if not isNoop:
-    ffiBody.add(bodyNode)
+
+  # Lift the body into an async impl registered in the per-library
+  # `ffiTeardownHook`, which the FFI thread awaits at shutdown (see ffi_thread.nim
+  # and ffiTeardownHook's docstring). The C wrapper no longer runs the body.
+  let teardownImplName = genSym(nskProc, "ffiTeardownImpl")
+  let teardownRegistration =
+    if isNoop:
+      newEmptyNode()
+    else:
+      quote:
+        proc `teardownImplName`(lib: ptr `libTypeName`): Future[void] {.async.} =
+          let `libParamName` = lib[]
+          `bodyNode`
+
+        ffiTeardownHook[`libTypeName`]() = `teardownImplName`
 
   let poolIdent = ident($libTypeName & "FFIPool")
   ffiBody.add quote do:
@@ -1690,7 +1724,7 @@ macro ffiDtor*(args: varargs[untyped]): untyped =
     when not declared(`poolIdent`):
       var `poolIdent`: FFIContextPool[`libTypeName`]
 
-  let stmts = newStmtList(poolDecl, ffiProc)
+  let stmts = newStmtList(teardownRegistration, poolDecl, ffiProc)
 
   when defined(ffiDumpMacros):
     echo stmts.repr
diff --git a/tests/unit/test_ffi_teardown.nim b/tests/unit/test_ffi_teardown.nim
new file mode 100644
index 0000000..7778791
--- /dev/null
+++ b/tests/unit/test_ffi_teardown.nim
@@ -0,0 +1,86 @@
+import std/[atomics, os]
+import unittest2
+import results
+import ffi
+
+# Exercises the async `{.ffiDtor.}` teardown hook: a non-empty dtor body is
+# lifted into an `ffiTeardown` overload the FFI thread awaits on its own event
+# loop right before it exits. `destroyFFIContext` must block until that
+# teardown finishes.
+
+type TeardownLib = object
+
+# declareLibrary emits an importc `lib<name>NimMain`; this test links as a plain
+# exe, so stub it (mirrors test_ffi_context.nim).
+{.emit: "void libteardownlibNimMain(void) {}".}
+
+declareLibrary("teardownlib", TeardownLib)
+
+var gTeardownRan: Atomic[bool]
+var gTeardownThreadId: Atomic[int]
+
+type NoopConfig {.ffi.} = object
+  dummy: int
+
+proc teardownlib_create*(
+    config: NoopConfig
+): Future[Result[TeardownLib, string]] {.ffiCtor.} =
+  return ok(TeardownLib())
+
+proc teardownlib_destroy*(lib: TeardownLib): Future[void] {.ffiDtor.} =
+  ## Async teardown: sleeps on the FFI event loop, then records that it ran and
+  ## on which thread. If destroy didn't wait, `gTeardownRan` would still be false
+  ## when the test checks it.
+  await sleepAsync(200.milliseconds)
+  gTeardownThreadId.store(getThreadId())
+  gTeardownRan.store(true)
+
+proc noopCallback(
+    retCode: cint, msg: ptr cchar, len: csize_t, userData: pointer
+) {.cdecl, gcsafe, raises: [].} =
+  discard
+
+proc encodedPtr(bytes: var seq[byte]): ptr byte =
+  if bytes.len == 0:
+    nil
+  else:
+    cast[ptr byte](addr bytes[0])
+
+proc createCtxWithLib(): ptr FFIContext[TeardownLib] =
+  ## Spins up a context via the ctor and waits until `myLib` is populated (set on
+  ## the FFI thread when the ctor request is dispatched), so teardown has a lib.
+  var cfg = cborEncode(TeardownlibCreateCtorReq(config: NoopConfig(dummy: 0)))
+  let ret = teardownlib_create(encodedPtr(cfg), cfg.len.csize_t, noopCallback, nil)
+  if ret.isNil():
+    return nil
+  let ctx = cast[ptr FFIContext[TeardownLib]](ret)
+  var tries = 0
+  while ctx[].myLib.isNil() and tries < 500:
+    os.sleep(5)
+    inc tries
+  ctx
+
+suite "async {.ffiDtor.} teardown hook":
+  test "destroy blocks until the async teardown body completes":
+    let ctx = createCtxWithLib()
+    check not ctx.isNil()
+    check not ctx[].myLib.isNil()
+
+    gTeardownRan.store(false)
+    gTeardownThreadId.store(0)
+    let callerTid = getThreadId()
+
+    check TeardownlibFFIPool.destroyFFIContext(ctx).isOk()
+
+    # If destroy returned before awaiting the teardown, this would be false.
+    check gTeardownRan.load()
+    # And it must have run on the FFI thread, not the caller's.
+    check gTeardownThreadId.load() != 0
+    check gTeardownThreadId.load() != callerTid
+
+  test "teardown runs exactly once per context":
+    gTeardownRan.store(false)
+    let ctx = createCtxWithLib()
+    check not ctx.isNil()
+    check TeardownlibFFIPool.destroyFFIContext(ctx).isOk()
+    check gTeardownRan.load()