> For clean Markdown of any page, append .md to the page URL. > For a complete documentation index, see https://docs.nvidia.com/openshell/llms.txt. > For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.nvidia.com/openshell/_mcp/server. # Sandbox Compute Drivers > Reference for Docker, Podman, MicroVM, Kubernetes, and Windows MXC sandbox compute drivers. The gateway's configured compute driver determines how OpenShell creates each sandbox. The CLI workflow stays the same across drivers: you create, connect to, inspect, stop, start, and delete sandboxes through the gateway API. Most compute drivers run the OpenShell supervisor inside the sandbox workload. The supervisor launches the agent process, applies policy, routes egress through the proxy, injects configured credentials, and maintains the gateway session. A driver may instead set `driver_reports_runtime_readiness`. In that mode, driver-reported readiness does not require a supervisor session. The canonical create-time policy is part of `DriverSandboxSpec`; a driver that enforces policy outside the standard supervisor fetches later revisions through the existing sandbox configuration API. The Windows MXC driver reports its own readiness. Stop stops compute but retains the sandbox record and the driver's persistent workspace boundary. Start reactivates the same driver resource. Delete remains independent and removes compute plus driver-owned persistent state. While a sandbox is stopped, gateway access paths and exposed services remain unavailable. Restarting the gateway preserves this intent. The gateway does not stop Docker or Podman containers during shutdown. At startup it sends idempotent start requests for Docker, Podman, and MicroVM sandboxes that were intended to run; already-running resources are unchanged, retained stopped compute is restarted, and explicitly stopped sandboxes remain stopped. Kubernetes workloads continue running independently of the gateway process. The gateway forwards one exact, persisted main-process specification to every driver. Drivers serialize that specification in `OPENSHELL_MAIN_PROCESS_SPEC`; they do not install an idle `sleep` workload or reconstruct argv with shell parsing. Runtime restart policies are disabled so an exited canonical process remains a terminal sandbox result. Exit code zero produces `Completed`; a nonzero or signal-normalized exit produces `Error` with the exact exit code. Driver and supervisor failures remain `Error`. ## Build with Selected Compute Drivers Source builds of `openshell-gateway` can include any subset of the Docker, Podman, Kubernetes, VM, and MXC drivers. Enable the corresponding `compute-driver-docker`, `compute-driver-podman`, `compute-driver-kubernetes`, `compute-driver-vm`, or `compute-driver-mxc` Cargo features. For example, build a Docker-only gateway with telemetry support: ```shell cargo build --release -p openshell-gateway --no-default-features --features telemetry,compute-driver-docker ``` On Windows, select only MXC with: ```shell cargo build --release -p openshell-gateway --no-default-features --features telemetry,compute-driver-mxc,bundled-z3 ``` The default `in-tree-compute-drivers` feature retains the full platform driver set, including MXC on Windows. MXC links only on Windows. The other four features link drivers on non-Windows platforms; on Windows they install registrations that report the driver as unsupported. A build with no default features and no driver features connects to external drivers only. Auto-detection probes only compiled registrations. To select a driver omitted from a custom build, configure its external `socket_path` as described below. ## Configure a Compute Driver Configure the compute driver on the gateway. Current releases accept one driver per gateway. Set the singular `compute_driver` key in the gateway TOML file: ```toml [openshell.gateway] compute_driver = "docker" ``` Reserved built-in values are `docker`, `podman`, `kubernetes`, `vm`, and `mxc`. The `mxc` driver is available only in native Windows gateway builds. Non-reserved names select an extension driver and require a `socket_path` in `[openshell.drivers.]`. When `compute_driver` is unset, the gateway auto-detects Kubernetes, then Podman, then Docker. Docker must respond on a known API socket. Podman first probes known API sockets and then asks the `podman` CLI for the active native or machine-backed socket. The VM driver is never auto-detected; configure it explicitly with `compute_driver = "vm"` or set `OPENSHELL_COMPUTE_DRIVER=vm` in the launch environment. `compute_driver` accepts exactly one scalar driver name. The legacy `compute_drivers` list is rejected by schema version 2. Common gateway options: | Gateway TOML option | Description | | ----------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------- | | `compute_driver = ""` | Select the compute driver. Built-in values are `docker`, `podman`, `kubernetes`, and `vm`; custom names require `[openshell.drivers.].socket_path`. | Set driver-specific values such as sandbox images, callback endpoints, network names, and VM sizing in the gateway TOML file. A TLS-enabled gateway-managed Docker, Podman, or VM driver requires a complete `guest_tls_ca`, `guest_tls_cert`, and `guest_tls_key` bundle in `[openshell.gateway]`; package-managed local TLS supplies it automatically. Driver tables reject those gateway-owned fields. Kubernetes projects guest TLS through a Secret instead. See the [Gateway Configuration File](./gateway-config) reference for the full schema and migration steps. Extension drivers use the same `compute_driver.proto` gRPC surface as the managed VM driver. For an out-of-tree driver, choose a driver name and point the gateway at the Unix socket the operator has already provisioned: ```toml [openshell.gateway] compute_driver = "kyma" [openshell.drivers.kyma] socket_path = "/run/openshell/kyma.sock" ``` For a launch-time socket override, pass the selected driver name with the socket path. The endpoint replaces normal driver construction for that name, including canonical built-in names: ```shell openshell-gateway --compute-driver kyma --compute-driver-socket /run/openshell/kyma.sock openshell-gateway --compute-driver docker --compute-driver-socket /run/openshell/docker.sock ``` The gateway connects to the operator-provided endpoint; it does not provision or supervise the remote driver. The operator must protect the socket so only the gateway uid can access it. Sandbox create supports `--cpu` and `--memory` for per-sandbox compute sizing. Docker and Podman apply them as runtime limits. Kubernetes applies them as both container requests and limits. The VM driver accepts the fields but currently ignores them. Sandbox create also accepts experimental driver-owned config through `--driver-config-json`. The value is a JSON object keyed by driver name. The gateway forwards only the block for the active driver, so a Kubernetes gateway receives the `kubernetes` object from a value such as: Nested keys inside each driver block use snake\_case. The top-level envelope keys are driver names, such as `kubernetes`, and are not part of the nested schema. ```shell openshell sandbox create \ --driver-config-json '{"kubernetes":{"pod":{"runtime_class_name":"kata-containers","priority_class_name":"batch-low"}}}' \ -- claude ``` Driver config is for fields without a stable public flag. Prefer `--cpu`, `--memory`, and `--gpu` for supported resource intent. When `--gpu` is present without a count, OpenShell treats it as a request for one GPU. Pass `--gpu COUNT` when requesting more than one GPU. Kubernetes maps the GPU count to the `nvidia.com/gpu` pod resource limit. Docker and Podman satisfy count-only GPU requests by selecting the requested number of NVIDIA CDI devices from the local CDI inventory in round-robin order. The drivers refresh the CDI inventory before validating or creating the sandbox, so CDI devices added or removed after gateway start can affect later creates. On WSL2 all-only runtimes, Docker or Podman can use `nvidia.com/gpu=all` as a compatibility fallback, where it counts as one selectable device. Exact GPU device selection remains driver-owned and requires `--gpu`. Docker and Podman accept `cdi_devices` as opaque CDI device names; replace the top-level `docker` key with `podman` when using the Podman driver, for example `{"docker":{"cdi_devices":["nvidia.com/gpu=0"]}}`. Explicit CDI device lists must not contain duplicates, and their length must match the effective GPU count. A single exact CDI device is compatible with the default `--gpu` request. The VM driver accepts `gpu_device_ids`, for example `{"vm":{"gpu_device_ids":["0000:2d:00.0"]}}`; the current VM implementation accepts at most one entry and allows either `--gpu` or `--gpu 1` when `gpu_device_ids` is set. ## Resource Capability Reporting API clients can read each configured driver's static resource request support from the protected `GetGatewayInfo` response. The capability response describes which request forms a driver implements. It does not report live resource inventory, availability, or scheduling capacity, so a supported request can still fail when the selected runtime cannot provide the resource. CPU and memory capabilities report whether the driver enforces a resource limit. GPU capabilities report whether the driver accepts a default GPU request and an explicit GPU count. An omitted `resource_capabilities` message, or an omitted CPU, memory, or GPU capability, means the driver did not report that capability. A reported `false` value means that request form is unsupported. For Kubernetes, `pod.runtime_class_name` maps to PodSpec `runtimeClassName`. It overrides the gateway's configured default runtime class for that sandbox, while a typed `SandboxTemplate.runtime_class_name` value from the API still takes precedence. Docker and Podman report the address through which their sandboxes can reach the gateway. If the primary listener covers that address, the gateway reuses it and sandbox JWT authentication restricts the supervisor to its callback RPC allowlist. If the primary listener is not reachable through that address, the gateway creates an additional callback-only listener. Use the primary endpoint for CLI, administrator, health, reflection, provider management, and HTTP requests. A `PermissionDenied` response from an additional callback-only listener is expected for those requests. Do not broaden the primary listener to `0.0.0.0` solely to make sandbox callbacks reachable. ## Docker Driver [Docker](https://www.docker.com/get-started/)-backed sandboxes run as containers on the gateway host. Use Docker for local development, single-machine gateways, and hosts that already use Docker Desktop or Docker Engine. The gateway talks to the Docker daemon to create sandbox containers. Docker Desktop and compatible macOS runtimes route `host.openshell.internal` through an IPv4 host-gateway alias. The gateway reuses an IPv4 primary listener that already covers loopback. Otherwise, the Docker driver requests a separate `127.0.0.1:` callback-only listener. For maintainer-level implementation details, refer to the [Docker driver README](https://github.com/NVIDIA/OpenShell/blob/main/crates/openshell-driver-docker/README.md). Select Docker with `compute_driver = "docker"` in `[openshell.gateway]`. Configure Docker driver values such as `socket_path`, `grpc_endpoint`, `network_name`, `sandbox_label`, `supervisor_bin`, `supervisor_image`, `image_pull_policy`, `ssh_socket_path`, and `sandbox_pids_limit` in `[openshell.drivers.docker]`. When `socket_path` is unset, the driver uses the same responsive local socket selected by auto-detection. An explicitly selected Docker driver falls back to `/var/run/docker.sock` when no candidate responds. When operating `openshell-driver-docker` as an external driver, set `OPENSHELL_OTLP_ENDPOINT` to export its spans. The driver continues W3C trace context from gateway RPCs and reports as `openshell-driver-docker`. Stop stops the existing Docker container without removing its writable layer or attached volumes. Start starts that same container. A durably stopped container stays stopped across gateway restart, and delete remains responsible for removing it. Graceful gateway shutdown stops running-intent Docker containers through the driver RPC without recording an explicit sandbox stop. On startup, the gateway reconciles that retained intent with an idempotent start request. Explicitly stopped sandboxes remain stopped. For GPU-backed Docker sandboxes, configure Docker CDI before starting the gateway so OpenShell can detect the daemon capability. ### Docker Driver Config Mounts Docker driver config accepts user-supplied `volume` and `tmpfs` mounts. It also accepts `bind` mounts when `[openshell.drivers.docker]` sets `enable_bind_mounts = true` in `gateway.toml`. See Docker's [storage documentation](https://docs.docker.com/engine/storage/) for more information. Docker local-driver named volumes created with bind options also expose gateway-host paths, so OpenShell treats them like bind mounts and requires `enable_bind_mounts = true`. Use a `volume` mount for existing Docker named volumes: ```shell docker volume create openshell-work openshell sandbox create \ --driver-config-json '{"docker":{"mounts":[{"type":"volume","source":"openshell-work","target":"/sandbox/work","read_only":false}]}}' \ -- claude ``` Bind mounts share gateway-host filesystem resources with the sandbox. They may be considered insecure because they can negate OpenShell controls such as workspace isolation and filesystem policy. Use them only when you understand and accept that loss of isolation. Use a `bind` mount only after enabling it in the Docker driver table: ```toml [openshell.drivers.docker] enable_bind_mounts = true ``` ```shell openshell sandbox create \ --driver-config-json '{"docker":{"mounts":[{"type":"bind","source":"/srv/openshell/work","target":"/sandbox/work","read_only":false}]}}' \ -- claude ``` Docker mount schema: | Type | Fields | | -------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | `bind` | `source`, `target`, optional `read_only` (`true` by default), optional `selinux_label` (`shared` for `:z` or `private` for `:Z`). `source` must be an absolute host path. Requires `enable_bind_mounts = true`. | | `volume` | `source`, `target`, optional `read_only` (`true` by default), optional `subpath`. The named volume must already exist. Docker local-driver bind-backed volumes require `enable_bind_mounts = true`. | | `tmpfs` | `target`, optional `options`, optional `size_bytes`, optional `mode`. | OpenShell rejects mount `source`, `target`, and Docker volume `subpath` values with surrounding whitespace. OpenShell also rejects mount targets that replace the workspace root or container root, or contain or are contained by the configured SSH socket or reserved `/opt/openshell`, `/etc/openshell`, `/etc/openshell-tls`, `/run/openshell`, `/run/openshell-sidecar`, and network namespace roots. These checks do not make host bind mounts safe. ## Podman Driver [Podman](https://podman.io/)-backed sandboxes run as rootless containers on the gateway host. Use Podman for Linux workstation workflows that avoid a rootful Docker daemon. The gateway talks to the Podman API socket. The Podman driver requires Podman 5.x, cgroups v2, rootless networking, and an active Podman user socket. When `socket_path` is not set, the driver probes known socket paths, then uses the `podman` CLI to resolve the active native or machine-backed connection. It fails to start if neither method finds a socket. For maintainer-level implementation details, refer to the [Podman driver README](https://github.com/NVIDIA/OpenShell/blob/main/crates/openshell-driver-podman/README.md) and [Podman networking notes](https://github.com/NVIDIA/OpenShell/blob/main/crates/openshell-driver-podman/NETWORKING.md). Select Podman with `compute_driver = "podman"` in `[openshell.gateway]`. Configure Podman driver values such as `socket_path`, `network_name`, `supervisor_image`, `stop_timeout_secs`, `image_pull_policy`, `grpc_endpoint`, `host_gateway_ip`, `ssh_socket_path`, and `sandbox_pids_limit` in `[openshell.drivers.podman]`. ### macOS Podman Socket Path On macOS, Homebrew-installed Podman does not create the default socket path that the driver probes (`~/.local/share/containers/podman/machine/podman.sock`). The actual API socket lives under `/var/folders/` in a path that macOS can rotate after a reboot. If the gateway fails with `Podman socket not found; is podman machine running?` while `podman machine list` shows a running machine, set the `OPENSHELL_PODMAN_SOCKET` environment variable to the dynamic socket path: ```shell export OPENSHELL_PODMAN_SOCKET="$(podman machine inspect --format '{{.ConnectionInfo.PodmanSocket.Path}}')" ``` Add this to your shell profile or gateway launch environment so it resolves correctly after each reboot. Alternatively, set `socket_path` in `[openshell.drivers.podman]` to the current path, but note that the path may change when macOS rotates `/var/folders/`. Podman sandboxes default to a 45-second graceful stop window before Podman escalates from `SIGTERM` to `SIGKILL`. Set `stop_timeout_secs` in gateway config, or `OPENSHELL_STOP_TIMEOUT` for the standalone driver, when a local runtime needs a different teardown window. Stop stops the existing Podman container while retaining its named workspace volume and driver-owned secrets. Start starts the same container. Delete is the operation that removes the container and named volume. Graceful gateway shutdown stops running-intent Podman containers through the driver RPC without recording an explicit sandbox stop. On startup, the gateway reconciles that retained intent with an idempotent start request while leaving explicitly stopped sandboxes alone. For proxy-required networks, the Podman driver also accepts the corporate egress proxy keys `https_proxy`, `no_proxy`, `proxy_auth_file`, `proxy_auth_allow_insecure`, and `proxy_connect_by_hostname`. The supervisor chains policy-approved TLS tunnels through the proxy with HTTP CONNECT instead of dialing destinations directly. See the [Gateway Configuration File](./gateway-config) reference for the full contract, including the cleartext-credential acknowledgement and the validated-IP CONNECT behavior. Podman preserves its runtime-selected AppArmor profile when `app_armor_profile` is omitted. Set `Unconfined` explicitly only when the supervisor's mount setup requires it. Explicit `RuntimeDefault` and `Localhost/` selections fail startup when Podman reports that AppArmor is unavailable. On macOS with `podman machine`, the driver uses gvproxy's host-loopback IP, `192.168.127.254`, for sandbox host aliases by default. Set `host_gateway_ip` only when your Podman machine uses a non-standard host-loopback address. On Linux, an empty `host_gateway_ip` keeps Podman's `host-gateway` resolver behavior. Direct local callbacks from rootless Podman require Podman to report the pasta network helper. Slirp4netns, other helpers, and Podman versions that do not report their helper require an explicitly remote `grpc_endpoint`; otherwise the gateway fails startup rather than leaving sandbox callbacks unreachable. Rootful Podman continues to use the configured network's bridge gateway address. ### Podman Driver Config Mounts Podman driver config accepts user-supplied `volume`, `tmpfs`, and `image` mounts. It also accepts `bind` mounts when `[openshell.drivers.podman]` sets `enable_bind_mounts = true` in `gateway.toml`. Podman local-driver named volumes created with bind options also expose gateway-host paths, so OpenShell treats them like bind mounts and requires `enable_bind_mounts = true`. Host bind mounts expose gateway host paths to sandbox requests, so they are disabled by default. Use a `volume` mount for existing Podman named volumes: ```shell podman volume create openshell-work openshell sandbox create \ --driver-config-json '{"podman":{"mounts":[{"type":"volume","source":"openshell-work","target":"/sandbox/work","read_only":false}]}}' \ -- claude ``` Bind mounts share gateway-host filesystem resources with the sandbox. They may be considered insecure because they can negate OpenShell controls such as workspace isolation and filesystem policy. Use them only when you understand and accept that loss of isolation. Use a `bind` mount only after enabling it in the Podman driver table: ```toml [openshell.drivers.podman] enable_bind_mounts = true ``` ```shell openshell sandbox create \ --driver-config-json '{"podman":{"mounts":[{"type":"bind","source":"/srv/openshell/work","target":"/sandbox/work","read_only":false}]}}' \ -- claude ``` Podman mount schema: | Type | Fields | | -------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | `bind` | `source`, `target`, optional `read_only` (`true` by default), optional `selinux_label` (`shared` for `:z` or `private` for `:Z`). `source` must be an absolute host path. Requires `enable_bind_mounts = true`. | | `volume` | `source`, `target`, optional `read_only` (`true` by default). The named volume must already exist. Podman local-driver bind-backed volumes require `enable_bind_mounts = true`. | | `tmpfs` | `target`, optional `options`, optional `size_bytes`, optional `mode`. | | `image` | `source`, `target`, optional `read_only` (`true` by default). | Podman `volume` and `image` mounts do not support `subpath` in OpenShell driver config, and OpenShell rejects `subpath` for those mount types. OpenShell rejects mount `source` and `target` values with surrounding whitespace. OpenShell also rejects mount targets that replace the workspace root, container root, supervisor files, `/etc/openshell`, `/etc/openshell-tls`, authentication material, or network namespace paths. These checks do not make host bind mounts safe. ## MicroVM Driver MicroVM-backed sandboxes run inside VM-backed isolation instead of a container boundary. Use MicroVM when workloads need a VM boundary instead of a local container boundary. The gateway uses the VM compute driver to create VM-backed sandboxes. MicroVM requires host virtualization support. It uses [libkrun](https://github.com/containers/libkrun) with Apple's [Hypervisor framework](https://developer.apple.com/documentation/hypervisor) on macOS, KVM on Linux, and [QEMU](https://www.qemu.org/) for GPU-backed sandboxes on Linux. The VM driver boots a cached immutable bootstrap ext4 root disk. When the requested sandbox image differs from the bootstrap image, the driver stages the registry image as an OCI layout, unpacks it inside a bootstrap VM with `umoci`, and caches the prepared image disk by image identity. Each sandbox receives that prepared disk read-only plus its own writable `overlay.ext4` disk for `/`, including `/sandbox` writes and runtime TLS material. The overlay persists for the sandbox lifetime and is deleted with the sandbox state directory. VM sandbox creation follows the same progress model as Kubernetes-backed sandboxes. The gateway accepts the sandbox, then the VM driver publishes watch events while it resolves the image, prepares or reuses the bootstrap and prepared image caches, creates the writable overlay, and starts the VM launcher. On graceful gateway shutdown, the gateway stops running-intent VMs through the driver RPC while retaining their launch records and writable overlays. On restart, the gateway starts a fresh VM driver process and reconciles the retained intent through the same idempotent start request used by the local container drivers. Running-intent VMs restart with their existing `overlay.ext4`, while explicitly stopped VMs remain stopped. Stopped VM state directories contain a marker that prevents startup from launching the VM. The driver retains `sandbox.pb`, `overlay.ext4`, and extension state, then removes the marker and restores the same overlay on start. For maintainer-level implementation details, refer to the [VM driver README](https://github.com/NVIDIA/OpenShell/blob/main/crates/openshell-driver-vm/README.md). ### Enable the VM Driver The VM driver is opt-in. Release packages can install `openshell-driver-vm`, but the gateway does not select it unless you configure the driver explicitly. Enable VM by setting `compute_driver = "vm"` in the gateway TOML file: ```toml [openshell.gateway] compute_driver = "vm" ``` For a launch-time override, set `OPENSHELL_COMPUTE_DRIVER=vm` in the gateway environment and restart the service. Configure VM driver values such as `grpc_endpoint`, `driver_dir`, `state_dir`, `default_image`, `bootstrap_image`, `vcpus`, `mem_mib`, `overlay_disk_mib`, and `krun_log_level` in `[openshell.drivers.vm]`. The VM `state_dir` stores overlay disks, console logs, runtime state, image-rootfs cache, and the private `run/compute-driver.sock` socket. The VM socket path is managed by the gateway and is not configurable through remote endpoint settings. The gateway starts `openshell-driver-vm` over a private Unix socket and passes its process ID so the driver can reject unexpected local clients. The driver's standalone TCP listener is disabled unless `--allow-unauthenticated-tcp` is set for local development. Scripts that invoke the experimental standalone driver directly must use `--grpc-endpoint` and the `--upstream-proxy*` option family. Schema v2 removes the previous `--openshell-endpoint`, `--https-proxy`, `--no-proxy`, and `--proxy-*` spellings; gateway-managed deployments do not use those options directly. ### Local image resolution The VM driver resolves sandbox images from a local container engine before falling back to registry pulls. It tries Docker first, then uses the same Podman socket discovery as the Podman driver. On Linux with Podman, enable the API socket so the driver can find local images: ```shell systemctl --user start podman.socket ``` ### Host Firewall The VM driver creates nftables rules on the host for each sandbox VM's TAP network interface. These rules provide NAT for VM connectivity and defense-in-depth isolation: unsolicited inbound connections to the VM are dropped, and the VM can only reach the gateway port on the host. Primary security enforcement (proxy-only egress and bypass detection) is handled by the sandbox supervisor inside the VM guest. On hosts with restrictive firewalls (e.g. firewalld), the host firewall may additionally block VM traffic that the driver's rules accept. If VM sandboxes cannot reach the network, verify that the host firewall allows forwarding and input for `vmtap-*` interfaces. See the [VM driver README](https://github.com/NVIDIA/OpenShell/blob/main/crates/openshell-driver-vm/README.md#host-side-nftables-rules) for details. ### Corporate Proxy Egress For proxy-required networks, the VM driver accepts the same corporate egress proxy keys as the Podman driver: `https_proxy`, `no_proxy`, `proxy_auth_file`, `proxy_auth_allow_insecure`, `proxy_connect_by_hostname`, and `proxy_ca_bundle`. The in-guest supervisor chains policy-approved TLS tunnels through the proxy with HTTP CONNECT instead of dialing destinations directly. The settings reach the guest supervisor on its command line through a per-sandbox argument file the driver writes into the overlay upperdir on every launch, so a sandbox cannot select, alter, or disable the proxy from inside the guest — including through image `ENV`, the sandbox environment, or files baked into the image at the paths the driver uses. A proxy on the corporate network needs no special address and works on every VM sandbox. The guest's callback to the gateway never traverses the proxy. A proxy on the gateway host itself works only for libkrun-backed (non-GPU) sandboxes, whose egress leaves through gvproxy: configure `https_proxy = "http://host.openshell.internal:"` rather than a `127.0.0.1` URL, because gvproxy NATs that alias to the host's `127.0.0.1`. GPU sandboxes use the QEMU/TAP backend, where `host.openshell.internal` resolves to the TAP host address and the driver's [host firewall rules](#host-firewall) allow the guest to reach only the gateway port on the host. The driver rejects a gateway-host proxy URL when a sandbox launches on QEMU instead of letting every CONNECT time out, so give GPU sandboxes a proxy address routable from the guest's masqueraded egress. Because a microVM has no bind mounts or container secrets, the driver stages the credential (root-only) and the CA bundle into the per-sandbox overlay disk and removes them with the sandbox. See the [Gateway Configuration File](./gateway-config) reference for the full contract, including the cleartext-credential acknowledgement and the validated-IP CONNECT behavior. ## Kubernetes Driver Kubernetes-backed sandboxes run as pods in the configured sandbox namespace. Use Kubernetes for shared clusters, remote compute, GPU scheduling, and operator-managed environments. Kubernetes workspace namespaces are an administrative trust boundary. In shared and managed modes, only the OpenShell gateway and its trusted Agent Sandbox controller may administer Sandbox CRs, sandbox pods, or the configured sandbox ServiceAccount in those namespaces. In operator mode, allowlist only namespaces where the platform operator preserves that exclusive control. Untrusted principals must not be able to create sandbox pods with fabricated owner references or use the sandbox ServiceAccount. The operator namespace allowlist is a trust grant, not a tenant isolation mechanism. Helm deployments set Kubernetes driver values through the chart. Canonical TOML places `namespace`, `service_account_name`, and `enable_user_namespaces` in `[openshell.drivers.kubernetes]`; schema version 2 rejects their historical `[openshell.gateway]` locations. For maintainer-level implementation details, refer to the [Kubernetes driver README](https://github.com/NVIDIA/OpenShell/blob/main/crates/openshell-driver-kubernetes/README.md). | Gateway configuration | Helm value | Description | | ------------------------------------------------------- | ------------------------------------------------------ | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | `compute_driver = "kubernetes"` | Not applicable | Select the Kubernetes compute driver. | | `[openshell.drivers.kubernetes].namespace` | `server.sandboxNamespace` | Set the namespace for sandbox resources. The Helm chart defaults to the release namespace when left empty. | | `[openshell.drivers.kubernetes].service_account_name` | `sandboxServiceAccount.name` | Set the Kubernetes service account assigned to sandbox pods and accepted by the Kubernetes driver's TokenReview bootstrap path. The Helm chart creates a dedicated sandbox service account by default. | | `[openshell.drivers.kubernetes].enable_user_namespaces` | `server.enableUserNamespaces` | Enable Kubernetes user namespaces for sandbox pods. | | `default_image` | `server.sandboxImage` | Set the default sandbox image. | | `image_pull_policy` | `server.sandboxImagePullPolicy` | Set the canonical sandbox pull policy: `always`, `if_not_present`, or `never`. `newer` is Podman-only. | | `image_pull_secrets` | `server.sandboxImagePullSecrets` | Attach Kubernetes image-pull Secrets to sandbox pods. Managed mode copies these explicitly named Secrets from the configured source namespace into each workspace namespace. In shared and operator modes, the Secrets must already exist in the sandbox namespace. | | `[managed_ssh_ingress]` | `networkPolicy.enabled` | In managed mode, create an SSH ingress policy in every workspace namespace. Helm configures the gateway namespace and pod selector automatically. Operator mode leaves namespace policy management to the platform operator. | | `grpc_endpoint` | `server.grpcEndpoint` | Set the gateway callback endpoint reachable from sandbox pods. Raw TOML and the standalone Kubernetes driver require an explicit endpoint because the sandbox namespace does not identify the gateway Service. Helm derives it from the release's gateway Service when the value is empty. | | `client_tls_secret_name` | `server.tls.clientTlsSecretName` | Mount sandbox client TLS materials from a Kubernetes secret. | | `supervisor_image` | `supervisor.image.repository` / `supervisor.image.tag` | Override the supervisor image that provides the `openshell-sandbox` binary. The default repository with an empty tag uses the version-pinned image built into the gateway. Changing the repository uses the effective gateway image tag, while setting a tag pins that version explicitly. | | `supervisor_image_pull_policy` | `supervisor.image.pullPolicy` | Set the canonical supervisor pull policy: `always`, `if_not_present`, or `never`. `newer` is Podman-only. | | `supervisor_sideload_method` | `supervisor.sideloadMethod` | How the supervisor binary is delivered into sandbox pods. Leave empty to auto-detect from cluster version. Set to `image-volume` to mount the supervisor OCI image directly as a volume (requires Kubernetes 1.33+ with the ImageVolume feature gate; GA in 1.36), or `init-container` to copy it through an init container on older clusters. | | `topology` | `supervisor.topology` | Set `combined` for the default single supervisor path, or `sidecar` to move pod-level network enforcement and the gateway session into a dedicated sidecar. | | `https_proxy` | `upstreamProxy.url` | Set the operator-owned `http://host:port` corporate forward proxy used for policy-approved TLS CONNECT egress. | | `no_proxy` | `upstreamProxy.noProxy` | Set destinations that bypass only the corporate proxy. OpenShell policy evaluation still applies. | | `proxy_auth_secret_name` | `upstreamProxy.authSecret.name` | Set the existing Secret name in the sandbox namespace that contains the proxy credential. Requires `sidecar` topology. | | `proxy_auth_secret_key` | `upstreamProxy.authSecret.key` | Set the Secret key containing the `user:pass` credential. Requires `sidecar` topology. | | `proxy_auth_allow_insecure` | `upstreamProxy.authAllowInsecure` | Set `true` to acknowledge that Basic authentication to an HTTP proxy is cleartext. Required with a proxy credential Secret. | | `proxy_connect_by_hostname` | `upstreamProxy.connectByHostname` | Send hostnames rather than validated IPs in CONNECT requests. Use only when proxy ACLs require hostname targets. | | `sidecar.proxy_uid` | `supervisor.sidecar.proxyUid` | Dedicated UID of at least `1000` used by the relaxed sidecar when process/binary-aware network policy is disabled. It must not match the workload UID. The default binary-aware sidecar runs as UID 0. The network init container exempts the effective sidecar UID from proxy redirection. | | `sidecar.process_binary_aware_network_policy` | `supervisor.sidecar.processBinaryAwareNetworkPolicy` | Keep process/binary-aware network policy enabled in `sidecar` topology. The default runs the sidecar as UID 0 with `SYS_PTRACE` and `DAC_READ_SEARCH`. Set false to run as `proxy_uid`, drop both capabilities, and enforce endpoint/L7 policy without matching `policy.binaries`. | | `app_armor_profile` | `server.appArmorProfile` | Set the sandbox agent container's AppArmor profile. Helm defaults this to `Unconfined` so AppArmor-enabled nodes do not block supervisor network namespace setup. Set the Helm value to an empty string to omit the field, or use `RuntimeDefault` or `Localhost/` for operator-managed profiles. | | `workspace_default_storage_size` | `server.workspaceDefaultStorageSize` | Set the default workspace PVC size for new sandboxes. | | `workspace_storage_class` | `server.workspaceStorageClass` | Set the `StorageClass` for the workspace PVC. Empty (default) omits `storageClassName` and uses the cluster's default `StorageClass`. Set this on clusters with no default `StorageClass`, otherwise the workspace PVC stays `Pending` and the sandbox never starts. | | `sa_token_ttl_secs` | `server.sandboxJwt.k8sSaTokenTtlSecs` | Set the projected ServiceAccount token TTL used for the bootstrap token exchange. | Managed-mode Secret copying requires the gateway ServiceAccount to create Secrets. Kubernetes RBAC cannot restrict Secret `create` by resource name, so the Helm chart grants cluster-wide Secret `create`; Secret `get` and `patch` remain limited to the explicitly configured TLS and image-pull Secret names. The driver creates copies only in gateway-owned managed namespaces. Do not reuse the gateway ServiceAccount for unrelated workloads. In `combined` topology, the agent container carries the Linux capabilities needed by the supervisor for network namespace setup, Landlock filesystem policy, process privilege changes, and network policy enforcement. In `sidecar` topology, the agent container runs as the resolved sandbox UID/GID with no added Linux capabilities. A root init container performs the nftables setup, and the long-running binary-aware sidecar runs as UID 0, drops default capabilities, and adds `SYS_PTRACE` plus `DAC_READ_SEARCH` for workload process identity resolution through shared `/proc`. The `sidecar.process_binary_aware_network_policy = false` setting runs it as the configured non-root `proxy_uid`, removes both capabilities, and relaxes network policy to endpoint/L7 matching only. The network sidecar owns gateway authentication and writes local policy/provider state to the process supervisor over a local control socket, so the agent container does not mount the sandbox bootstrap token or client TLS secret in the default sidecar path. The provider environment is refreshed by the network sidecar after settings polls and streamed to the process supervisor so future child processes can see updated provider env without gateway access in the agent container. Sidecar mode keeps gateway session and SSH behavior. The process supervisor applies Landlock filesystem policy and child seccomp filters where supported, but it does not perform root-to-sandbox privilege dropping or supervisor identity mount isolation. Network policy still runs in the sidecar, and sidecar pods set `shareProcessNamespace: true` so the network sidecar can resolve process/binary identity through `/proc/`. The Kubernetes driver creates namespaced `agents.x-k8s.io` `Sandbox` resources from the Kubernetes SIG Apps [agent-sandbox](https://github.com/kubernetes-sigs/agent-sandbox) project. It detects the served Sandbox API at runtime, caches the selected API version for the gateway process, and uses `v1beta1` when available before falling back to `v1alpha1`, so supported Agent Sandbox installations work without version-specific operator configuration. The Agent Sandbox controller turns those resources into sandbox pods and related storage. Stop patches the existing resource rather than deleting it. For `v1beta1`, the driver sets `spec.operatingMode` to `Suspended` or `Running`. For `v1alpha1`, it sets `spec.replicas` to `0` or `1`. The Sandbox resource and its workspace PVC keep their identity across both operations. Stop returns only after the controller reports suspension and deletes the old pod, so an immediate start cannot race the prior pod's termination. If Agent Sandbox is upgraded in place, restart the OpenShell gateway after the controller and CRD rollout completes so the gateway can detect the served API versions again. `Sandbox.spec.volumeClaimTemplates` is immutable after creation. To change storage configuration, delete the sandbox and create a new one with the updated spec. ### Kubernetes Driver Config PVC Mounts Kubernetes driver config can mount existing PersistentVolumeClaims into the agent container. Use this when storage is provisioned outside OpenShell and a sandbox should mount selected PVC subpaths instead of using the default OpenShell-created `/sandbox` workspace PVC. ```shell openshell sandbox create \ --driver-config-json '{ "kubernetes": { "volumes": [{ "name": "user-data", "persistent_volume_claim": { "claim_name": "pvc-user-data-123", "read_only": false } }], "containers": { "agent": { "volume_mounts": [ { "name": "user-data", "mount_path": "/sandbox/.openshell/workspace", "sub_path": "workspace", "read_only": false }, { "name": "user-data", "mount_path": "/sandbox/.openshell/memory", "sub_path": "memory", "read_only": false } ] } } } }' \ -- claude ``` Kubernetes PVC mount schema: | Field | Description | | ---------------------------------------------- | ------------------------------------------------------------------------------------------------- | | `volumes[].name` | Pod volume name. It must be a DNS-1123 label, unique, and not use OpenShell-managed volume names. | | `volumes[].persistent_volume_claim.claim_name` | Existing PVC name in the sandbox namespace. It must be a DNS-1123 subdomain name. | | `volumes[].persistent_volume_claim.read_only` | Optional. Defaults to `true`. Set `false` to allow read-write mounts. | | `containers.agent.volume_mounts[].name` | References a volume declared in `volumes`. | | `containers.agent.volume_mounts[].mount_path` | Absolute, normalized container path for the agent mount. | | `containers.agent.volume_mounts[].sub_path` | Optional relative PVC subpath. Absolute paths and `..` are rejected. | | `containers.agent.volume_mounts[].read_only` | Optional. Defaults to `true`. It cannot be `false` when the PVC volume is read-only. | OpenShell rejects duplicate volume names, mounts that reference unknown volumes, protected mount targets, and mounts that replace OpenShell TLS, supervisor, ServiceAccount token, or SPIFFE paths. Read-write PVC access requires `read_only: false` on both the PVC volume and each writable mount. Any driver-config mount under `/sandbox` disables the default `/sandbox` workspace PVC injection for that sandbox. Only the explicit mount paths persist through the external PVC; other `/sandbox` paths come from the current sandbox image. ## Sandbox User Identity The policy can set `process.run_as_user` and `process.run_as_group` independently. Each explicit field wins. The active compute driver supplies the identity for omitted fields. Explicit numeric values may use any non-root Linux UID/GID from `1` through `4294967294`. OpenShell rejects `0` as root and `4294967295` as the invalid identity sentinel. Low numeric identities can inherit permissions from matching accounts, files, volumes, or devices, so choose them with the same care as any other runtime identity. ### Docker / Podman Docker and Podman inspect the final image and use its OCI `USER` declaration as a per-field fallback. Supported forms include `app`, `app:staff`, a numeric UID whose passwd entry supplies its primary GID, and an accountless numeric pair such as `1234:1235`. The driver pins container creation to the immutable image ID it inspected. The supervisor validates any required names inside that image and preserves the declared name or numeric components for both direct and SSH children. When `USER` omits the group, the supervisor uses the user's numeric primary GID. It does not modify `/etc/passwd` or `/etc/group`. Docker also inspects OCI `WorkingDir`. An absolute value becomes the agent workspace; an empty, root (`/`), or explicit `/sandbox` value uses the managed `/sandbox` compatibility workspace. OpenShell creates and owns that compatibility workspace. Any other workdir must already exist in the immutable image without symlink components. The completed UID/GID and supplementary groups must already be able to traverse every parent and write and enter the workdir. OpenShell does not change that directory's ownership or mode. A one-shot validator drops to that identity and uses kernel effective-access checks, including POSIX ACL grants and LSM denials. It rejects workdirs that overlap the OCI runtime namespaces under `/proc`, `/sys`, or `/dev`, and rejects overlap with actual OpenShell control paths. Docker checks the original image filesystem in the final supervisor and rejects image `VOLUME` declarations that would mask the workdir or one of its parents before validation. The resolved workspace is the cwd and `HOME` for direct and SSH children. The supervisor itself starts from `/`, so a missing or invalid workspace is handled during readiness instead of preventing the container runtime from starting it. Sandbox creation fails before readiness if a required `USER` component is missing, malformed, unknown, ambiguous, or resolves to UID/GID 0. An image without `USER` therefore works only when policy explicitly provides both identity fields. ### Kubernetes / OpenShift The Kubernetes driver auto-detects the sandbox UID from OpenShift SCC namespace annotations: * `openshift.io/sa.scc.uid-range` (format: `/`, e.g. `1000000000/10000`) provides the UID. * `openshift.io/sa.scc.supplemental-groups` provides the GID when present; otherwise the resolved UID is used as the GID. * On non-OpenShift clusters, or when annotations are absent, the driver falls back to `1000`. You can override autodetection with explicit `sandbox_uid` / `sandbox_gid` config in `[openshell.drivers.kubernetes]`. When set, the driver skips namespace annotation lookup entirely. The resolved UID/GID appear in: * Supervisor container environment variables (`OPENSHELL_SANDBOX_UID`, `OPENSHELL_SANDBOX_GID`) for direct kernel-level privilege dropping without `/etc/passwd` lookups. * PVC init container `securityContext.runAsUser/runAsGroup/fsGroup` for workspace ownership operations. ### VM Driver The VM driver preserves an image-provided `sandbox` account when `sandbox_uid` and `sandbox_gid` are omitted. Images without that account use UID/GID `1000`. Explicit values in `[openshell.drivers.vm]` override the image account. Persisted overlays retain the UID/GID recorded when they were created. An unmarked overlay recovers identity from concrete overlay or prepared-image state, an explicit override, or the current image; the driver never assigns legacy `10001:10001` without persisted evidence. ### Custom Images Docker and Podman custom images do not need a baked-in `"sandbox"` user. Declare a non-root OCI `USER`, or set both process identity fields explicitly in policy. Named image users require matching account entries; a numeric `UID:GID` pair does not. For Docker, declare an absolute OCI `WORKDIR` to select the workspace. Images with no working directory, `WORKDIR /`, or `WORKDIR /sandbox` use OpenShell's managed `/sandbox` compatibility workspace. For any other Docker path, create the directory in the image and grant the final process identity write and execute permission in the Dockerfile. Podman, Kubernetes/OpenShift, and VM sandboxes continue to use `/sandbox`.