diff --git a/Project.toml b/Project.toml
index 4644639..9a82eb1 100644
--- a/Project.toml
+++ b/Project.toml
@@ -8,7 +8,7 @@ authors = [
     "Jake Bolewski <clima-software@caltech.edu>",
     "Gabriele Bozzola <gbozzola@caltech.edu>",
 ]
-version = "0.6.4"
+version = "0.6.5"
 
 [weakdeps]
 CUDA = "052768ef-5323-5732-b1bb-66c8b64840ba"
diff --git a/ext/ClimaCommsMPIExt.jl b/ext/ClimaCommsMPIExt.jl
index d04732b..d75e958 100644
--- a/ext/ClimaCommsMPIExt.jl
+++ b/ext/ClimaCommsMPIExt.jl
@@ -1,10 +1,26 @@
 module ClimaCommsMPIExt
 
 import MPI
+import ClimaComms: mpicomm, set_mpicomm!
 import ClimaComms
 
-ClimaComms.MPICommsContext(device = ClimaComms.device()) =
-    ClimaComms.MPICommsContext(device, MPI.COMM_WORLD)
+const CLIMA_COMM_WORLD = Ref{typeof(MPI.COMM_WORLD)}()
+
+function set_mpicomm!()
+    CLIMA_COMM_WORLD[] = MPI.COMM_WORLD
+    return CLIMA_COMM_WORLD[]
+end
+
+ClimaComms.mpicomm(::ClimaComms.MPICommsContext) = CLIMA_COMM_WORLD[]
+
+# For backwards compatibility
+function Base.getproperty(ctx::ClimaComms.MPICommsContext, sym::Symbol)
+    if sym === :mpicomm
+        return ClimaComms.mpicomm(ctx)
+    else
+        return getfield(ctx, sym)
+    end
+end
 
 function ClimaComms.init(ctx::ClimaComms.MPICommsContext)
     if !MPI.Initialized()
@@ -19,9 +35,9 @@ function ClimaComms.init(ctx::ClimaComms.MPICommsContext)
         end
         # assign GPUs based on local rank
         local_comm = MPI.Comm_split_type(
-            ctx.mpicomm,
+            mpicomm(ctx),
             MPI.COMM_TYPE_SHARED,
-            MPI.Comm_rank(ctx.mpicomm),
+            MPI.Comm_rank(mpicomm(ctx)),
         )
         ClimaComms._assign_device(ctx.device, MPI.Comm_rank(local_comm))
         MPI.free(local_comm)
@@ -32,36 +48,36 @@ end
 ClimaComms.device(ctx::ClimaComms.MPICommsContext) = ctx.device
 
 ClimaComms.mypid(ctx::ClimaComms.MPICommsContext) =
-    MPI.Comm_rank(ctx.mpicomm) + 1
+    MPI.Comm_rank(mpicomm(ctx)) + 1
 ClimaComms.iamroot(ctx::ClimaComms.MPICommsContext) = ClimaComms.mypid(ctx) == 1
-ClimaComms.nprocs(ctx::ClimaComms.MPICommsContext) = MPI.Comm_size(ctx.mpicomm)
+ClimaComms.nprocs(ctx::ClimaComms.MPICommsContext) = MPI.Comm_size(mpicomm(ctx))
 
-ClimaComms.barrier(ctx::ClimaComms.MPICommsContext) = MPI.Barrier(ctx.mpicomm)
+ClimaComms.barrier(ctx::ClimaComms.MPICommsContext) = MPI.Barrier(mpicomm(ctx))
 
 ClimaComms.reduce(ctx::ClimaComms.MPICommsContext, val, op) =
-    MPI.Reduce(val, op, 0, ctx.mpicomm)
+    MPI.Reduce(val, op, 0, mpicomm(ctx))
 
 ClimaComms.reduce!(ctx::ClimaComms.MPICommsContext, sendbuf, recvbuf, op) =
-    MPI.Reduce!(sendbuf, recvbuf, op, ctx.mpicomm; root = 0)
+    MPI.Reduce!(sendbuf, recvbuf, op, mpicomm(ctx); root = 0)
 
 ClimaComms.reduce!(ctx::ClimaComms.MPICommsContext, sendrecvbuf, op) =
-    MPI.Reduce!(sendrecvbuf, op, ctx.mpicomm; root = 0)
+    MPI.Reduce!(sendrecvbuf, op, mpicomm(ctx); root = 0)
 
 ClimaComms.allreduce(ctx::ClimaComms.MPICommsContext, sendbuf, op) =
-    MPI.Allreduce(sendbuf, op, ctx.mpicomm)
+    MPI.Allreduce(sendbuf, op, mpicomm(ctx))
 
 ClimaComms.allreduce!(ctx::ClimaComms.MPICommsContext, sendbuf, recvbuf, op) =
-    MPI.Allreduce!(sendbuf, recvbuf, op, ctx.mpicomm)
+    MPI.Allreduce!(sendbuf, recvbuf, op, mpicomm(ctx))
 
 ClimaComms.allreduce!(ctx::ClimaComms.MPICommsContext, sendrecvbuf, op) =
-    MPI.Allreduce!(sendrecvbuf, op, ctx.mpicomm)
+    MPI.Allreduce!(sendrecvbuf, op, mpicomm(ctx))
 
 ClimaComms.bcast(ctx::ClimaComms.MPICommsContext, object) =
-    MPI.bcast(object, ctx.mpicomm; root = 0)
+    MPI.bcast(object, mpicomm(ctx); root = 0)
 
 function ClimaComms.gather(ctx::ClimaComms.MPICommsContext, array)
     dims = size(array)
-    lengths = MPI.Gather(dims[end], 0, ctx.mpicomm)
+    lengths = MPI.Gather(dims[end], 0, mpicomm(ctx))
     if ClimaComms.iamroot(ctx)
         dimsout = (dims[1:(end - 1)]..., sum(lengths))
         arrayout = similar(array, dimsout)
@@ -69,11 +85,11 @@ function ClimaComms.gather(ctx::ClimaComms.MPICommsContext, array)
     else
         recvbuf = nothing
     end
-    MPI.Gatherv!(array, recvbuf, 0, ctx.mpicomm)
+    MPI.Gatherv!(array, recvbuf, 0, mpicomm(ctx))
 end
 
 ClimaComms.abort(ctx::ClimaComms.MPICommsContext, status::Int) =
-    MPI.Abort(ctx.mpicomm, status)
+    MPI.Abort(mpicomm(ctx), status)
 
 
 # We could probably do something fancier here?
@@ -171,7 +187,7 @@ function graph_context(
         for n in 1:length(recv_bufs)
             MPI.Recv_init(
                 recv_bufs[n],
-                ctx.mpicomm,
+                mpicomm(ctx),
                 recv_reqs[n];
                 source = recv_ranks[n],
                 tag = tag,
@@ -181,7 +197,7 @@ function graph_context(
         for n in 1:length(send_bufs)
             MPI.Send_init(
                 send_bufs[n],
-                ctx.mpicomm,
+                mpicomm(ctx),
                 send_reqs[n];
                 dest = send_ranks[n],
                 tag = tag,
@@ -226,7 +242,7 @@ function ClimaComms.start(
             ghost.recv_bufs[n],
             ghost.recv_ranks[n],
             ghost.tag,
-            ghost.ctx.mpicomm,
+            mpicomm(ghost.ctx),
             ghost.recv_reqs[n],
         )
     end
@@ -236,7 +252,7 @@ function ClimaComms.start(
             ghost.send_bufs[n],
             ghost.send_ranks[n],
             ghost.tag,
-            ghost.ctx.mpicomm,
+            mpicomm(ghost.ctx),
             ghost.send_reqs[n],
         )
     end
@@ -254,9 +270,9 @@ function ClimaComms.progress(
     ghost::Union{MPISendRecvGraphContext, MPIPersistentSendRecvGraphContext},
 )
     if isdefined(MPI, :MPI_ANY_SOURCE) # < v0.20
-        MPI.Iprobe(MPI.MPI_ANY_SOURCE, ghost.tag, ghost.ctx.mpicomm)
+        MPI.Iprobe(MPI.MPI_ANY_SOURCE, ghost.tag, mpicomm(ghost.ctx))
     else # >= v0.20
-        MPI.Iprobe(MPI.ANY_SOURCE, ghost.tag, ghost.ctx.mpicomm)
+        MPI.Iprobe(MPI.ANY_SOURCE, ghost.tag, mpicomm(ghost.ctx))
     end
 end
 
diff --git a/src/mpi.jl b/src/mpi.jl
index 524cc2c..189c6fa 100644
--- a/src/mpi.jl
+++ b/src/mpi.jl
@@ -6,9 +6,15 @@
 A MPI communications context, used for distributed runs.
 [`AbstractCPUDevice`](@ref) and [`CUDADevice`](@ref) device options are currently supported.
 """
-struct MPICommsContext{D <: AbstractDevice, C} <: AbstractCommsContext
+struct MPICommsContext{D <: AbstractDevice} <: AbstractCommsContext
     device::D
-    mpicomm::C
+    function MPICommsContext(dev::AbstractDevice = device())
+        set_mpicomm!()
+        return new{typeof(dev)}(dev)
+    end
 end
 
 function MPICommsContext end
+
+function set_mpicomm! end
+function mpicomm end