SmolBox

Run Python, JavaScript, and other programs in disposable or retained microVMs from Elixir.

SmolBox is an Elixir client and supervised execution runtime for smolvm, which runs lightweight virtual machines on your own hosts. smolvm provides the machines and worker API; SmolBox tracks commands, results, and cleanup from your application's supervision tree.

Use it when an Elixir application needs to call a Python library, run a JavaScript processing step, or execute a script in a separate guest environment. You operate the workers and prepare images containing the languages and dependencies you need. The default execution API runs one command in its own disposable VM. SmolBox.Machines also manages retained machines that can run successive commands and expose TCP services through fixed worker-host port mappings. Startup workloads and console diagnostics add immutable entrypoint, command, environment and working directory on managed image machines. Console snapshots and bounded streaming aid boot diagnosis; application stdout/stderr and automatic restart policies remain unsupported by qualified upstream behavior.

Interactive terminal sessions add bounded input/output, resizing and durable exit-or-uncertainty tracking on retained machines.

Long-running commands and background launch support longer builds and persistent services, with explicit budgets and typed launch evidence.

Guest paths and larger files authorize project and home directories and buffered files up to 16 MiB. Defaults remain /workspace and 1 MiB. Policies persist through controller recovery and apply to ordinary command working directories too.

Images and registry artifacts let retained machines provision from approved, digest-pinned registry artifacts or OCI images. Preparation identity survives controller recovery. OCI machines support managed image pulls with typed results and conservative handling of lost responses. Shared host caches remain an operator responsibility.

Try the community workspace app for a browser walkthrough of SmolBox 0.2.0: persistent machines, commands, a real terminal, larger file transfers, background launch evidence and a mapped service, with PostgreSQL-backed recovery.

Physical Linux provisioning measurements compare fresh machines, exports, checkpoints, branches and reuse, including preparation, resource use and verified cleanup.

Why use SmolBox?

Executing a command is only part of integrating a worker. Your application also needs to handle duplicate requests, lost responses, restarts, and leftover VMs. SmolBox provides that execution lifecycle:

• Keep one identity for a request. Submitting the same scoped ID and specification returns the existing handle. A different specification under that ID is rejected.
• Track work across application restarts. A durable store adapter preserves execution records so the runtime can resume observation and cleanup. When a command may have run but its result was lost, SmolBox records an unknown outcome instead of automatically running it again.
• Manage files and machine lifecycle. The runtime stages input files, limits captured output and file sizes, and tracks cancellation and cleanup separately from the command's exit status.

What execution looks like

This example assumes a supervised runtime named MyApp.Sandboxes is already configured. python_artifact is the registered Python image's identity map ("id", "sha256", "architecture"), and profile is one of the runtime's allowed execution profiles. Getting started walks through the complete worker and runtime setup.

alias SmolBox.{Command, ExecutionSpec}

argv = ["python", "-c", "print(6 * 7)"]
{:ok, command} = Command.new(argv, timeout_secs: 10)

{:ok, spec} =
  ExecutionSpec.new(
    scope: "demo",
    id: "answer-001",
    artifact: python_artifact,
    profile: profile,
    command: command
  )

{:ok, handle} = SmolBox.submit(MyApp.Sandboxes, spec)
{:ok, ^handle} = SmolBox.submit(MyApp.Sandboxes, spec)

{:ok, execution} =
  SmolBox.await(MyApp.Sandboxes, handle, 90_000)

%{state: :completed, result: result} = execution
%{exit_code: 0, stdout: stdout} = result

IO.write(stdout)
# Prints: 42

The second submission reuses the execution; it does not start another command. The matches above show a successful run. In your application, handle errors, nonzero exits, and unknown outcomes explicitly. Cleanup can still be pending when await/3 returns, and its wait timeout does not cancel the command. Keep the runtime supervised and inspect the record with SmolBox.fetch/3; see result handling and troubleshooting.

If your application already owns machine lifecycle and persistence, SmolBox.Client exposes individual worker operations for creating machines, executing commands, streaming output, and transferring files.

Installation and first run

SmolBox 0.3.0 adds registry provisioning, managed exports, checkpoint capture and independent restore, and live branches. It keeps smolvm 1.19.0 as the default, unchanged from 0.2.1. See the feature and platform evidence and physical Linux comparisons.

Upgrading from 0.2.x requires coordination before using the new features. Upgrade every shared controller, reader and store adapter before writing the new records. The PostgreSQL example needs no additional SQL migration, but older releases cannot read the new formats, including retained history after deletion. Read Upgrading to 0.3.0 for capabilities, accounting and rollback limits. Workers are installed separately; pin existing workers to their actual version.

Upgrading from 0.1.x requires coordination. Upgrade shared controllers and store adapters together, apply the example's machine/port migrations if using it, and upgrade workers separately or explicitly pin their installed versions. New record formats and retained tombstones constrain rollback. Start with Upgrading to 0.2.0.

Add SmolBox to your application's mix.exs:

{:smolbox, "~> 0.3.0"}

Run mix deps.get. The API documentation includes the guides below. A local checkout can instead be used with {:smolbox, path: "../smolbox"}.

To run the local walkthrough, you need:

• Elixir 1.18 or later, using a tested Elixir/OTP pair.
• A dedicated worker on Linux x86_64 with KVM or macOS Apple Silicon: smolvm 1.19.0 by default, or explicitly configured 1.17.0, 1.16.1, 1.16.0, 1.14.6 or 1.14.1 workers.
• The host's resize2fs tool for 1.14.6, 1.16.0, 1.16.1, 1.17.0 and 1.19.0 disk requests below template sizes. On macOS, install e2fsprogs; see the runtime prerequisites.
• A prepared Python image for that worker's architecture, with its SHA-256 recorded. The guide links to the image preparation commands and host capacity requirements.

Follow Getting started to configure supervision, stage a Python file, submit it, read its output file, and confirm cleanup. The walkthrough uses an in-memory store and needs no database. Applications that need restart recovery must provide a durable SmolBox.Store adapter; a complete PostgreSQL host example is included in the repository.

Managed persistent machines

SmolBox.Machines gives a machine its own durable identity, ownership, resource reservation, and create/start/stop/delete lifecycle. Run successive commands with independent execution identities while retaining guest files. With a durable store, a new controller reconnects to the same machine after an application restart. Machines remain until explicitly deleted; uncertain work blocks reuse without replay.

See Managed persistent machines for the API, recovery procedures, the required coordinated store upgrade, and a two-process PostgreSQL walkthrough. The existing disposable API remains supported.

Managed branches

Branch an idle, offline bare guest into independent running children on the same smolvm 1.19.0 worker. Memory and disks are copied, held release is explicit, and durable dependencies and backing capacity survive controller restarts.

Checkpoint execution

Managed idle, offline bare guests can also capture checkpoints and restore independent machines on smolvm 1.19.0, with durable history and explicit retention.

SmolBox can restore an operator-approved idle, offline checkpoint into a separate disposable machine for each execution. See Executing from a checkpoint for approval, examples and record schema v3 upgrade requirements.

Current scope

SmolBox has real execution and durable recovery tests on Linux x86_64 and macOS Apple Silicon with smolvm 1.14.1, 1.14.6, 1.16.0, 1.16.1, 1.17.0 and 1.19.0; results are recorded in Compatibility. SmolBox 0.3.0 and 0.2.1 default to 1.19.0 and 0.2.0 to 1.17.0, while 0.1.4–0.1.5 default to 1.16.1; 0.1.3 defaults to 1.16.0 and 0.1.2 to 1.14.6. Before adopting the 1.19.0 default with an older worker, explicitly configure runtime_version: "1.17.0", "1.16.1", "1.16.0", "1.14.6" or "1.14.1", or follow the worker upgrade procedure. The package does not upgrade an external worker. A version mismatch prevents new execution; arbitrary upstream releases and automatic fallback are not accepted. Published versions 0.1.4 and 0.1.5 select 1.16.1 by default after qualification and a cleanup change that separates disposal from preservation. A failed graceful stop can still leave unknown work retained; see the recovery rules.

SmolBox supports explicit outbound hostname/CIDR policies with smolvm 1.16.0, 1.16.1, 1.17.0 and 1.19.0. Offline remains the default. See Controlled network access for setup and validation boundaries.

SmolBox relies on smolvm's isolation model for running untrusted code. Your deployment must protect worker access and configure host resource limits, networking, and credentials. A subsequent validation campaign tested these controls and failure recovery in one constrained Linux deployment; see Linux deployment validation for its results and limits. The walkthrough does not configure that deployment.

The library's supported qualification remains :development. Its allocation settings and admission reservations do not enforce host quotas; unsupported hard-control options remain rejected. Read Deployment boundaries when configuring your workers.

Worker installation, runtime image preparation, authentication, and host storage remain application/operator responsibilities. SmolBox runs programs already available in the prepared image. TypeScript needs a compiler or runner in that image; SmolBox does not build or publish user functions.

Guides

The engineering blog has practical articles about running programs from Elixir and managing their execution lifecycle.

Contributing

Source: hfiguera/smolbox. Maintainer toolchain versions are in .tool-versions; tested consumer combinations are recorded in the compatibility guide. From this repository, mix ci runs deterministic checks without contacting a real worker. Generate this site with MIX_ENV=dev mix docs --warnings-as-errors. Live worker tests are a separate opt-in operation described in the repository's CI guide.