Skip to content

With the sandbox enabled, the model gets a sandbox tool that runs any shell command (sh -c) in a fresh container, without approval. The container is the security boundary: inside it anything goes, including package installs and, unless limited, network access. Coding agents are not run in it; where they may work without asking is their confinement.

The only host path a container sees is the conversation’s workspace, data_dir/sandbox/<conversation ID>, mounted at /workspace and kept between calls. Everything else in the container is discarded when the command exits. No host environment variables are passed in. Files the container writes may be owned by its user (often root) on Linux hosts. Mob refuses to write through a link a command planted in the workspace.

One section, reloaded live. Disabling it withdraws the tool from later turns.

sandbox:
enabled: true
runtime: "" # containerd, apple or wslc; empty: the platform's own
image: python:3.13-slim
network: none # empty: the runtime's default network
cpus: 2 # whole CPUs per command; 0: the runtime's default
memory: 1g # at least 64m; empty: the runtime's default
isolation: gvisor # Linux with containerd only; empty: the runtime's own

MOB_SANDBOX_ENABLED, MOB_SANDBOX_RUNTIME, MOB_SANDBOX_IMAGE, MOB_SANDBOX_NETWORK, MOB_SANDBOX_CPUS, MOB_SANDBOX_MEMORY and MOB_SANDBOX_ISOLATION override it.

apps/mob/src/sandbox/container.ts drives each platform’s own runtime through its CLI, so Mob needs no Docker Desktop and gains no native dependencies:

Platform Runtime Isolation
Linux containerd through nerdctl, in its own mob namespace shares the host kernel; with rootful containerd an escape is root. Prefer rootless, or gVisor
macOS Apple’s container a VM per container
Windows WSL containers (wslc) a VM per session

The three share Docker’s run flags, so one argument builder serves them. A Runner seam lets tests run without any runtime installed.

Requirements: containerd needs nerdctl and CNI plugins beside it (plain ctr is not supported). Apple’s runtime needs macOS 26 on Apple silicon and container system start. wslc needs a WSL release that includes it. When the runtime is not ready the service logs it and the tool returns the reason.

  • A command runs for 2 minutes by default, at most 5, and never past its turn. On timeout or cancellation the named container is force-removed; on containerd Mob runs stop -t 0 first, since under runsc rm -f alone can leave it Exited.
  • Stdout and stderr are each kept up to 1 MiB; how much the model reads is decided by the large tool output policy.
  • A nonzero exit is a result the model reads, not a failure.
  • The image is pulled on first use. A failed pull does not block a run while the runtime answers, so a local image still works offline.
  • network, cpus and memory become --network none, --cpus N and --memory <MiB>m, spelled so all three CLIs read them alike. CPUs are whole because Apple’s CLI takes an integer. Memory requires a unit, since the CLIs disagree on a bare number. Network accepts only none; named networks differ per runtime.
  • With network: none the tool’s description tells the model there is no network. DNS and every interface but loopback are gone, so package installs fail; use an image with the tools already in it. Allowing only some hosts is not offered.
  • What a limit bounds follows the runtime: a cgroup on containerd and wslc, the container’s VM on Apple’s runtime, where memory also covers its kernel and an unset value means its configured default (1 GiB unless changed). On Linux, CPU and memory limits need the cgroup controllers available to containerd; rootless containerd needs them delegated.
  • There is no process-count limit: only nerdctl documents --pids-limit.

isolation: gvisor puts gVisor’s user-space kernel between the container and the host’s, without a VM. It requires the containerd runtime; on macOS and Windows, or with another runtime, the configuration is refused rather than failing on the first command.

  • nerdctl runs it as --runtime runsc, which runs runsc under containerd’s runc shim, so only runsc has to be on containerd’s PATH (install guide).
  • Network, CPU and memory limits and the workspace mount behave as under runc.
  • System-call-heavy work is slower. On a VM without KVM (gVisor’s systrap platform) container start-up was unchanged, while 2,000 file writes and cats took 13 s against 1.6 s under runc. Package installs and builds feel this most.
  • Rootless containerd with runsc is untested.

mob container uses the same runtime directly, with --network, --cpus, --memory and --isolation, so a user can confirm it works before enabling the sandbox:

Terminal window
mob container status # which runtime, and is it running
mob container pull python:3.13-slim
mob container run -v "$PWD:/work" -w /work alpine:3.20 ls -la

scripts/verify-sandbox.sh and scripts/verify-sandbox.ps1 check a machine end to end, including from inside a container that each limit took effect; see sandbox-verification.md.