diff --git a/pkgs/clan-cli/clan_cli/vms/inspect.py b/pkgs/clan-cli/clan_cli/vms/inspect.py
index a2f5dc513..ada68be7c 100644
--- a/pkgs/clan-cli/clan_cli/vms/inspect.py
+++ b/pkgs/clan-cli/clan_cli/vms/inspect.py
@@ -30,7 +30,6 @@ class VmConfig:
 
     waypipe: bool = False
 
-
     def __post_init__(self) -> None:
         if isinstance(self.flake_url, str):
             self.flake_url = FlakeId(self.flake_url)
diff --git a/pkgs/clan-cli/clan_cli/vms/run.py b/pkgs/clan-cli/clan_cli/vms/run.py
index 73dd2ecf2..30c5fe254 100644
--- a/pkgs/clan-cli/clan_cli/vms/run.py
+++ b/pkgs/clan-cli/clan_cli/vms/run.py
@@ -3,17 +3,23 @@ import importlib
 import json
 import logging
 import os
-from contextlib import ExitStack
+import socket
+import subprocess
+import time
+from collections.abc import Iterator
+from contextlib import ExitStack, contextmanager
 from pathlib import Path
 from tempfile import TemporaryDirectory
 
-from clan_cli.cmd import Log, run
+from clan_cli.cmd import CmdOut, Log, handle_output, run
 from clan_cli.completions import add_dynamic_completer, complete_machines
 from clan_cli.dirs import module_root, user_cache_dir, vm_state_dir
-from clan_cli.errors import ClanError
+from clan_cli.errors import ClanCmdError, ClanError
 from clan_cli.facts.generate import generate_facts
 from clan_cli.machines.machines import Machine
 from clan_cli.nix import nix_shell
+from clan_cli.qemu.qga import QgaSession
+from clan_cli.qemu.qmp import QEMUMonitorProtocol
 
 from .inspect import VmConfig, inspect_vm
 from .qemu import qemu_command
@@ -107,14 +113,94 @@ def prepare_disk(
     return disk_img
 
 
-def run_vm(
+@contextmanager
+def start_vm(
+    args: list[str],
+    packages: list[str],
+    extra_env: dict[str, str],
+    stdout: int | None = None,
+    stderr: int | None = None,
+) -> Iterator[subprocess.Popen]:
+    env = os.environ.copy()
+    env.update(extra_env)
+    cmd = nix_shell(packages, args)
+    with subprocess.Popen(cmd, env=env, stdout=stdout, stderr=stderr) as process:
+        try:
+            yield process
+        finally:
+            process.terminate()
+            try:
+                # Fix me: This should in future properly shutdown the VM using qmp
+                process.wait(timeout=5)
+            except subprocess.TimeoutExpired:
+                process.kill()
+
+
+class QemuVm:
+    def __init__(
+        self,
+        machine: Machine,
+        process: subprocess.Popen,
+    ) -> None:
+        self.machine = machine
+        self.process = process
+        self.state_dir = vm_state_dir(self.machine.flake, self.machine.name)
+        self.qmp_socket_file = self.state_dir / "qmp.sock"
+        self.qga_socket_file = self.state_dir / "qga.sock"
+
+    # wait for vm to be up then connect and return qmp instance
+    @contextmanager
+    def qmp_connect(self) -> Iterator[QEMUMonitorProtocol]:
+        with QEMUMonitorProtocol(
+            address=str(os.path.realpath(self.qmp_socket_file)),
+        ) as qmp:
+            qmp.connect()
+            yield qmp
+
+    @contextmanager
+    def qga_connect(self, timeout_sec: float = 100) -> Iterator[QgaSession]:
+        sock = socket.socket(socket.AF_UNIX, socket.SOCK_STREAM)
+        try:
+            # try to reconnect a couple of times if connection refused
+            socket_file = os.path.realpath(self.qga_socket_file)
+            start_time = time.time()
+            while time.time() - start_time < timeout_sec:
+                try:
+                    sock.connect(str(socket_file))
+                except ConnectionRefusedError:
+                    time.sleep(0.1)
+                else:
+                    break
+                sock.connect(str(socket_file))
+            yield QgaSession(sock)
+        finally:
+            sock.close()
+
+    def wait_up(self, timeout_sec: float = 60) -> None:
+        start_time = time.time()
+        while time.time() - start_time < timeout_sec:
+            if self.process.poll() is not None:
+                msg = "VM failed to start. Qemu process exited with code {self.process.returncode}"
+                raise ClanError(msg)
+            if self.qmp_socket_file.exists():
+                break
+            time.sleep(0.1)
+
+    def wait_down(self) -> int:
+        return self.process.wait()
+
+
+@contextmanager
+def spawn_vm(
     vm: VmConfig,
     *,
     cachedir: Path | None = None,
     socketdir: Path | None = None,
     nix_options: list[str] | None = None,
     portmap: list[tuple[int, int]] | None = None,
-) -> None:
+    stdout: int | None = None,
+    stderr: int | None = None,
+) -> Iterator[QemuVm]:
     if portmap is None:
         portmap = []
     if nix_options is None:
@@ -141,7 +227,7 @@ def run_vm(
         # TODO: We should get this from the vm argument
         nixos_config = build_vm(machine, cachedir, nix_options)
 
-        state_dir = vm_state_dir(str(vm.flake_url), machine.name)
+        state_dir = vm_state_dir(vm.flake_url, machine.name)
         state_dir.mkdir(parents=True, exist_ok=True)
 
         # specify socket files for qmp and qga
@@ -185,24 +271,71 @@ def run_vm(
 
         packages = ["nixpkgs#qemu"]
 
-        env = os.environ.copy()
+        extra_env = {}
         if vm.graphics and not vm.waypipe:
             packages.append("nixpkgs#virt-viewer")
             remote_viewer_mimetypes = module_root() / "vms" / "mimetypes"
-            env["XDG_DATA_DIRS"] = (
-                f"{remote_viewer_mimetypes}:{env.get('XDG_DATA_DIRS', '')}"
+            extra_env["XDG_DATA_DIRS"] = (
+                f"{remote_viewer_mimetypes}:{os.environ.get('XDG_DATA_DIRS', '')}"
             )
 
         with (
             start_waypipe(qemu_cmd.vsock_cid, f"[{vm.machine_name}] "),
             start_virtiofsd(virtiofsd_socket),
+            start_vm(
+                qemu_cmd.args, packages, extra_env, stdout=stdout, stderr=stderr
+            ) as process,
         ):
-            run(
-                nix_shell(packages, qemu_cmd.args),
-                env=env,
-                log=Log.BOTH,
-                error_msg=f"Could not start vm {machine}",
-            )
+            qemu_vm = QemuVm(machine, process)
+            qemu_vm.wait_up()
+
+            try:
+                yield qemu_vm
+            finally:
+                try:
+                    with qemu_vm.qmp_connect() as qmp:
+                        qmp.command("system_powerdown")
+                    qemu_vm.wait_down()
+                except OSError:
+                    pass
+                # TODO: add a timeout here instead of waiting indefinitely
+
+
+def run_vm(
+    vm_config: VmConfig,
+    *,
+    cachedir: Path | None = None,
+    socketdir: Path | None = None,
+    nix_options: list[str] | None = None,
+    portmap: list[tuple[int, int]] | None = None,
+) -> None:
+    with spawn_vm(
+        vm_config,
+        cachedir=cachedir,
+        socketdir=socketdir,
+        nix_options=nix_options,
+        portmap=portmap,
+        stdout=subprocess.PIPE,
+        stderr=subprocess.PIPE,
+    ) as vm:
+        stdout_buf, stderr_buf = handle_output(vm.process, Log.BOTH)
+        args: list[str] = vm.process.args  # type: ignore[assignment]
+        cmd_out = CmdOut(
+            stdout=stdout_buf,
+            stderr=stderr_buf,
+            cwd=Path.cwd(),
+            command_list=args,
+            returncode=vm.process.returncode,
+            msg=f"Could not start vm {vm_config.machine_name}",
+            env={},
+        )
+
+        if vm.process.returncode != 0:
+            raise ClanCmdError(cmd_out)
+        rc = vm.wait_down()
+        if rc != 0:
+            msg = f"VM exited with code {rc}"
+            raise ClanError(msg)
 
 
 def run_command(
diff --git a/pkgs/clan-cli/tests/helpers/vms.py b/pkgs/clan-cli/tests/helpers/vms.py
deleted file mode 100644
index 0b0247ea6..000000000
--- a/pkgs/clan-cli/tests/helpers/vms.py
+++ /dev/null
@@ -1,132 +0,0 @@
-import contextlib
-import os
-import socket
-import sys
-import threading
-import traceback
-from collections.abc import Iterator
-from pathlib import Path
-from time import sleep
-
-from clan_cli.dirs import vm_state_dir
-from clan_cli.errors import ClanError
-from clan_cli.qemu.qga import QgaSession
-from clan_cli.qemu.qmp import QEMUMonitorProtocol
-
-from . import cli
-
-
-def find_free_port() -> int:
-    """Find an unused localhost port from 1024-65535 and return it."""
-    with contextlib.closing(socket.socket(type=socket.SOCK_STREAM)) as sock:
-        sock.bind(("127.0.0.1", 0))
-        return sock.getsockname()[1]
-
-
-class VmThread(threading.Thread):
-    def __init__(self, machine_name: str, ssh_port: int | None = None) -> None:
-        super().__init__()
-        self.machine_name = machine_name
-        self.ssh_port = ssh_port
-        self.exception: Exception | None = None
-        self.daemon = True
-
-    def run(self) -> None:
-        try:
-            cli.run(
-                ["vms", "run", self.machine_name, "--publish", f"{self.ssh_port}:22"]
-            )
-        except Exception as ex:
-            # print exception details
-            print(traceback.format_exc(), file=sys.stderr)
-            print(sys.exc_info()[2], file=sys.stderr)
-            self.exception = ex
-
-
-def run_vm_in_thread(machine_name: str, ssh_port: int | None = None) -> VmThread:
-    # runs machine and prints exceptions
-    if ssh_port is None:
-        ssh_port = find_free_port()
-
-    vm_thread = VmThread(machine_name, ssh_port)
-    vm_thread.start()
-    return vm_thread
-
-
-# wait for qmp socket to exist
-def wait_vm_up(machine_name: str, vm: VmThread, flake_url: str | None = None) -> None:
-    if flake_url is None:
-        flake_url = str(Path.cwd())
-    socket_file = vm_state_dir(flake_url, machine_name) / "qmp.sock"
-    timeout: float = 600
-    while True:
-        if vm.exception:
-            msg = "VM failed to start"
-            raise ClanError(msg) from vm.exception
-        if timeout <= 0:
-            msg = f"qmp socket {socket_file} not found. Is the VM running?"
-            raise TimeoutError(msg)
-        if socket_file.exists():
-            break
-        sleep(0.1)
-        timeout -= 0.1
-
-
-# wait for vm to be down by checking if qmp socket is down
-def wait_vm_down(machine_name: str, vm: VmThread, flake_url: str | None = None) -> None:
-    if flake_url is None:
-        flake_url = str(Path.cwd())
-    socket_file = vm_state_dir(flake_url, machine_name) / "qmp.sock"
-    timeout: float = 300
-    while socket_file.exists():
-        if vm.exception:
-            msg = "VM failed to start"
-            raise ClanError(msg) from vm.exception
-        if timeout <= 0:
-            msg = f"qmp socket {socket_file} still exists. Is the VM down?"
-            raise TimeoutError(msg)
-        sleep(0.1)
-        timeout -= 0.1
-
-
-# wait for vm to be up then connect and return qmp instance
-@contextlib.contextmanager
-def qmp_connect(
-    machine_name: str, vm: VmThread, flake_url: str | None = None
-) -> Iterator[QEMUMonitorProtocol]:
-    if flake_url is None:
-        flake_url = str(Path.cwd())
-    state_dir = vm_state_dir(flake_url, machine_name)
-    wait_vm_up(machine_name, vm, flake_url)
-    with QEMUMonitorProtocol(
-        address=str(os.path.realpath(state_dir / "qmp.sock")),
-    ) as qmp:
-        qmp.connect()
-        yield qmp
-
-
-# wait for vm to be up then connect and return qga instance
-@contextlib.contextmanager
-def qga_connect(
-    machine_name: str, vm: VmThread, flake_url: str | None = None
-) -> Iterator[QgaSession]:
-    if flake_url is None:
-        flake_url = str(Path.cwd())
-    state_dir = vm_state_dir(flake_url, machine_name)
-    wait_vm_up(machine_name, vm, flake_url)
-
-    sock = socket.socket(socket.AF_UNIX, socket.SOCK_STREAM)
-    try:
-        # try to reconnect a couple of times if connection refused
-        socket_file = os.path.realpath(state_dir / "qga.sock")
-        for _ in range(100):
-            try:
-                sock.connect(str(socket_file))
-            except ConnectionRefusedError:
-                sleep(0.1)
-            else:
-                break
-            sock.connect(str(socket_file))
-        yield QgaSession(sock)
-    finally:
-        sock.close()
diff --git a/pkgs/clan-cli/tests/test_vars_deployment.py b/pkgs/clan-cli/tests/test_vars_deployment.py
index 22296909b..d9d127497 100644
--- a/pkgs/clan-cli/tests/test_vars_deployment.py
+++ b/pkgs/clan-cli/tests/test_vars_deployment.py
@@ -1,14 +1,17 @@
 import json
+from contextlib import ExitStack
 from pathlib import Path
 
 import pytest
 from age_keys import SopsSetup
 from clan_cli import cmd
+from clan_cli.clan_uri import FlakeId
+from clan_cli.machines.machines import Machine
 from clan_cli.nix import nix_eval, run
+from clan_cli.vms.run import inspect_vm, spawn_vm
 from fixtures_flakes import generate_flake
 from helpers import cli
 from helpers.nixos_config import nested_dict
-from helpers.vms import qga_connect, run_vm_in_thread, wait_vm_down
 from root import CLAN_CORE
 
 
@@ -61,9 +64,9 @@ def test_vm_deployment(
         flake_template=CLAN_CORE / "templates" / "minimal",
         machine_configs={"m1_machine": machine1_config, "m2_machine": machine2_config},
     )
-    monkeypatch.chdir(flake.path)
+
     sops_setup.init()
-    cli.run(["vars", "generate"])
+    cli.run(["vars", "generate", "--flake", str(flake.path)])
     # check sops secrets not empty
     for machine in ["m1_machine", "m2_machine"]:
         sops_secrets = json.loads(
@@ -94,13 +97,15 @@ def test_vm_deployment(
         ).stdout.strip()
         assert "no-such-path" not in shared_secret_path
     # run nix flake lock
-    cmd.run(["nix", "flake", "lock"])
-    vm_m1 = run_vm_in_thread("m1_machine")
-    vm_m2 = run_vm_in_thread("m2_machine")
-    with (
-        qga_connect("m1_machine", vm_m1) as qga_m1,
-        qga_connect("m2_machine", vm_m2) as qga_m2,
-    ):
+    cmd.run(["nix", "flake", "lock"], cwd=flake.path)
+
+    vm1_config = inspect_vm(machine=Machine("m1_machine", FlakeId(str(flake.path))))
+    vm2_config = inspect_vm(machine=Machine("m2_machine", FlakeId(str(flake.path))))
+    with ExitStack() as stack:
+        vm1 = stack.enter_context(spawn_vm(vm1_config))
+        vm2 = stack.enter_context(spawn_vm(vm2_config))
+        qga_m1 = stack.enter_context(vm1.qga_connect())
+        qga_m2 = stack.enter_context(vm2.qga_connect())
         # check my_secret is deployed
         _, out, _ = qga_m1.run(
             "cat /run/secrets/vars/m1_generator/my_secret", check=True
@@ -122,9 +127,3 @@ def test_vm_deployment(
             check=False,
         )
         assert returncode != 0
-        qga_m1.exec_cmd("poweroff")
-        qga_m2.exec_cmd("poweroff")
-        wait_vm_down("m1_machine", vm_m1)
-        wait_vm_down("m2_machine", vm_m2)
-    vm_m1.join()
-    vm_m2.join()
diff --git a/pkgs/clan-cli/tests/test_vms_cli.py b/pkgs/clan-cli/tests/test_vms_cli.py
index 7944fecf0..57d9faf7a 100644
--- a/pkgs/clan-cli/tests/test_vms_cli.py
+++ b/pkgs/clan-cli/tests/test_vms_cli.py
@@ -2,10 +2,12 @@ from pathlib import Path
 from typing import TYPE_CHECKING
 
 import pytest
+from clan_cli.clan_uri import FlakeId
+from clan_cli.machines.machines import Machine
+from clan_cli.vms.run import inspect_vm, spawn_vm
 from fixtures_flakes import FlakeForTest, generate_flake
 from helpers import cli
 from helpers.nixos_config import nested_dict
-from helpers.vms import qga_connect, qmp_connect, run_vm_in_thread, wait_vm_down
 from root import CLAN_CORE
 from stdout import CaptureOutput
 
@@ -84,30 +86,18 @@ def test_vm_persistence(
         machine_configs=config,
     )
 
-    monkeypatch.chdir(flake.path)
+    vm_config = inspect_vm(machine=Machine("my_machine", FlakeId(str(flake.path))))
 
-    vm = run_vm_in_thread("my_machine")
-
-    # wait for the VM to start and connect qga
-    with qga_connect("my_machine", vm) as qga:
+    with spawn_vm(vm_config) as vm, vm.qga_connect() as qga:
         # create state via qmp command instead of systemd service
         qga.run("echo 'dream2nix' > /var/my-state/root", check=True)
         qga.run("echo 'dream2nix' > /var/my-state/test", check=True)
         qga.run("chown test /var/my-state/test", check=True)
         qga.run("chown test /var/user-state", check=True)
         qga.run("touch /var/my-state/rebooting", check=True)
-        qga.exec_cmd("poweroff")
-
-        # wait for socket to be down (systemd service 'poweroff' rebooting machine)
-        wait_vm_down("my_machine", vm)
-
-    vm.join()
 
     ## start vm again
-    vm = run_vm_in_thread("my_machine")
-
-    ## connect second time
-    with qga_connect("my_machine", vm) as qga:
+    with spawn_vm(vm_config) as vm, vm.qga_connect() as qga:
         # check state exists
         qga.run("cat /var/my-state/test", check=True)
         # ensure root file is owned by root
@@ -131,7 +121,3 @@ def test_vm_persistence(
             "systemctl --failed | tee /tmp/yolo | grep -q '0 loaded units listed' || ( cat /tmp/yolo && false )"
         )
         assert exitcode == 0, out
-
-        with qmp_connect("my_machine", vm) as qmp:
-            qmp.command("system_powerdown")
-    vm.join()