smol — Python SDK

Embed isolated microVM sandboxes directly in your Python code. Same API locally (embedded engine, no server) or against the smolfleet cloud — the backend is chosen via ConnectOptions / SMOL_CLOUD_TOKEN. Mirrors the Node SDK.

Supported platforms (native local transport): macOS Apple Silicon, and Linux x64/arm64 with glibc ≥ 2.34 (RHEL 9, Ubuntu 22.04+, Debian 12, Amazon Linux 2023; the wheel is tagged manylinux_2_34). The cloud transport works anywhere the wheel installs. Not yet published: macOS Intel, and glibc < 2.34.

from smol import Machine, MachineConfig, ResourceSpec

# Local — no server. The SDK is in your process; the VMM is a separate,
# seccomp/Landlock-confined helper.
with Machine.create(MachineConfig(resources=ResourceSpec(cpus=2, memory_mb=1024, network=True))) as m:
    res = m.run("python:3.12", ["python", "-c", "print(2 ** 10)"])
    res.assert_success()
    print(res.stdout)            # 1024
    m.write_file("/tmp/in.txt", "hi")
    print(m.read_file("/tmp/in.txt").decode())

# Branch a prepared machine: a CoW clone of its RAM and disks, typically under
# 200ms, so a warm environment is reused instead of rebuilt. Pass network=True
# whenever an image has to be pulled.
golden = Machine.create(MachineConfig(image="alpine", network=True, forkable=True))
branch = golden.branch("b1")

# Cloud (smolfleet) — create() waits until it is ready for work.
from smol import ConnectOptions
m = Machine.create(
    MachineConfig(image="alpine:3.20"),
    ConnectOptions(target="cloud"),  # uses SMOL_CLOUD_TOKEN
)
try:
    print(m.exec(["echo", "ready for work"]).stdout)
finally:
    m.delete()

Async: AsyncMachine (non-blocking)

Machine is synchronous — each call blocks the calling thread. When you're driving many machines from one event loop (a fleet of disposable workers), use AsyncMachine: the same API, but every I/O method is a coroutine that runs off the loop, so launches and calls overlap instead of serializing.

import asyncio
from smol import AsyncMachine, MachineConfig, ConnectOptions, PortSpec

async def main():
    cfg = MachineConfig(image="alpine:3.20", ports=[PortSpec(host=8080, guest=8080)])
    conn = ConnectOptions(target="cloud")  # or SMOL_CLOUD_TOKEN
    # Launch a fleet concurrently — none blocks the loop.
    machines = await asyncio.gather(*(AsyncMachine.create(cfg, conn) for _ in range(8)))
    try:
        await asyncio.gather(*(m.wait_until_ready() for m in machines))
        # Reach a service inside a vm via the authed connect bridge (no tunnel):
        health = await machines[0].request(8080, "healthz")
    finally:
        await asyncio.gather(*(m.delete() for m in machines))

asyncio.run(main())

Every Machine method has an awaitable counterpart on AsyncMachine (create/connect/exec/wait_until_ready/request/fork/…), plus async with for auto-delete. endpoint(port) stays synchronous — it only builds a URL and does no I/O.

Fused multi-policy rollouts

RolloutClient is the thin generation boundary for TRL, Unsloth, and custom RL loops. The node keeps one vLLM engine hot, verifies immutable LoRA versions, and submits cross-policy cohorts concurrently so vLLM can continuously batch them.

from smol import RolloutClient

rollouts = RolloutClient("http://127.0.0.1:8080/api/v1", "qwen")
rollouts.ensure_vllm_executor(
    endpoint="http://127.0.0.1:8000",
    adapter_root="/var/lib/smol/adapters",
    fallback_pool="isolated-rollouts",
)
rollouts.publish_policy("experiment-a", "step-40", "/var/lib/smol/adapters/a-40")
result = rollouts.generate(
    idempotency_key="experiment-a-step-40-batch-7",
    policy="experiment-a",
    prompts=[[1, 2, 3]],
    max_tokens=64,
    temperature=0.9,
    logprobs=1,
)

Inside a forked rollout worker, no configuration is required: RolloutClient() discovers its authenticated node assignment from /etc/smolvm/fork-env and automatically groups workers from the same fork batch into a bounded cohort. Pass auto_fork_cohort=False only when the application already supplies an explicit cohort_id, cohort_size, and cohort_max_wait_ms.

Optional framework adapters are explicit imports, so the base SDK remains free of PyTorch, PEFT, Transformers, Unsloth, and vLLM dependencies:

from smol.integrations import (
    UnslothVllmExecutor,
    add_transformers_forkpoint,
    publish_peft_adapter,
)

The vLLM backend must bind to loopback, enable runtime LoRA updates, and reserve one spare CPU LoRA slot so a new version can load before the old version drains.

NeMo Gym sandbox provider

Install the optional integration and select the same smol provider for local SmolVM or Smol Cloud:

pip install 'smolmachines[nemo-gym]'

sandbox:
  smol:
    target: local                 # or cloud; cloud reuses `smol auth login`
    checkpoints:
      ghcr.io/acme/swe:ready:
        machine: swe-golden       # running MachineConfig(forkable=True) machine
        ports: [8000]
        resources:                # describe the prepared golden's capacity
          cpu: 4
          memory_mib: 8192
          disk_gib: 20
        provider_options:         # describe its inherited egress policy
          allow_hosts: [api.example.com]
    fork_batch_window_ms: 2        # coalesce concurrent episode creates
    fork_batch_size: 32
  default_metadata:
    sandbox-api: smol

NeMo Gym discovers the provider through its standard nemo_gym.sandbox_providers entry point. A normal image creates a fresh microVM. An exact image match in checkpoints instead creates every episode as a live RAM/disk copy-on-write fork of that prepared machine, so repositories, dependencies, services, and caches can already be running when the agent takes its first action. The provider implements exec, files, resource limits, scoped egress, declared service ports, entrypoint overrides, TTL cleanup, and the same configuration for local and cloud targets. Handles serialize without credentials and reconnect through the receiving SDK session, as required by DeepSWE's agent/verifier lifecycle. Cloud checkpoint episodes with a TTL use Smol Cloud's durable lease controller, so they are reclaimed even if the NeMo Gym process exits unexpectedly.

Checkpoint forks inherit the golden's resource shape, network policy, and running workload. Declare those inherited properties in the checkpoint mapping. A task's resource request may be smaller than the declared capacity, while its network policy, entrypoint, and ports must match exactly; incompatible requests fail instead of silently running with different isolation. NeMo Gym's standard sandbox provider contract supports episode-from-golden forks.

To branch an arbitrary live trajectory state, create that source as branchable and call the Smol provider's extension at the decision point:

from nemo_gym.sandbox.providers.base import SandboxSpec
from smol.nemo_gym import SmolProvider

provider = SmolProvider(target="local")  # or target="cloud"
checkpoint = await provider.create(
    SandboxSpec(
        image="ghcr.io/acme/swe:ready",
        provider_options={"branchable": True},
    )
)
await provider.exec(checkpoint, "./agent-step-1")
await provider.exec(checkpoint, "./agent-step-2")

branches = await provider.branch(checkpoint, count=16, name_prefix="candidate")

The first fan-out waits for active commands and freezes the source at that exact RAM/filesystem/process state. Every returned sandbox is an independent COW leaf; the frozen source may create more siblings from the same state but cannot execute more commands. Close the branches before the source. A configured checkpoint fork is already a leaf, so it cannot request another live branch until SmolVM supports nested fork generations.

Harbor environment provider

Run Harbor or Terminal-Bench trials as isolated copy-on-write SmolVM forks:

pip install 'smolmachines[harbor]'

harbor run \
  --dataset terminal-bench@2.0 \
  --agent oracle \
  --env smol.harbor:SmolEnvironment \
  --n-concurrent 16

The provider uses Harbor's standard custom-environment interface and works with the same SDK configuration locally or on Smol Cloud. By default, the first trial for each distinct environment creates a warm checkpoint from its published docker_image; concurrent and later trials fork clean RAM/disk COW clones from that checkpoint. Set --environment-kwarg auto_checkpoint=false for cold one-machine-per-trial behavior.

To use an environment prepared before the Harbor job, map its image reference or Harbor environment hash to the running checkpoint:

environment:
  import_path: smol.harbor:SmolEnvironment
  kwargs:
    target: cloud
    checkpoints:
      ghcr.io/acme/swe:ready:
        machine: mach-prepared-swe
        resources:
          cpus: 4
          memory_mb: 8192
          storage_mb: 20480
        network_mode: public

Checkpoint forks preserve initialized process and filesystem state rather than merely reusing image layers. The initial implementation supports Linux single-container tasks with a published docker_image; Docker Compose and Dockerfile-only tasks fail clearly instead of silently changing semantics.

Architecture

• Pure-Python layer (python/smol): Machine, transports, types, errors — zero third-party deps (the cloud transport uses only urllib).
• Native core (src/lib.rs, crate smol-py): a pyo3 extension that drives the smolvm engine for the local path — the Python analogue of the smol-node NAPI crate. The local API is synchronous (the engine blocks). The extension is in your process; the VMM is not — it runs as a separate smol-vmm helper, seccomp- and Landlock-confined on Linux.
• Cloud transport: a REST client to smolfleet /v1 whose request/response shapes match smolfleet's OpenAPI contract (Bearer smk_…).

Disposable workers: wait for ready, then connect (cloud)

Launching a machine as a disposable agent runtime has two easy-to-miss steps; both are first-class here.

Machine.create() already waits for the machine to be ready — not merely started. state == "started" means the VM process launched; the guest is still booting and is not usable yet. Acting on started is the classic teardown race (works on a slow cold start, times out on a warm one). Gate on the unambiguous signal:

m = Machine.create(
    MachineConfig(image="alpine:3.20", ports=[PortSpec(host=8080, guest=8080)]),
    ConnectOptions(target="cloud"),
)
try:
    # create() already waited: the guest agent is reachable and the published
    # port is accepting connections.
    # Reach a service INSIDE the vm through the authenticated connect bridge —
    # no Cloudflare/localhost.run tunnel, no public exposure, no egress allow-list.
    # Have the worker LISTEN on a published port and connect *inbound*:
    print(m.request(8080, "healthz").decode())     # authed HTTP to the guest port
    ep = m.endpoint(8080, "/socket")               # or build a ws:// url for your ws client
    # websocket.connect(ep.ws_url, additional_headers=ep.headers)
finally:
    m.delete()

# Machine.connect() intentionally does not wait for readiness:
existing = Machine.connect(machine_id, ConnectOptions(target="cloud"))
existing.wait_until_ready()

API

• Machine (sync) / AsyncMachine (awaitable, non-blocking) — identical surface; see the async example above.
• RolloutClient — publish versioned LoRAs and generate single- or multi-policy cohorts.
• Machine.create(config=None, conn=None) — create and start a machine; cloud waits for ready is True before returning.
• Machine.connect(machine_id, conn=None) — attach without waiting; call wait_until_ready() before use.
• machine.exec(command, opts=None) / machine.run(image, command, opts=None) → ExecResult
• machine.read_file(path) → bytes / machine.write_file(path, data, mode=None)
• machine.ready() / machine.ready_at() / machine.wait_until_ready(timeout_s=120, interval_s=1) (cloud)
• machine.endpoint(port, path=None) → PortEndpoint / machine.request(port, path=None, method="GET", data=None) → bytes (cloud connect bridge)
• machine.pull_image(image) / machine.list_images() (local)
• machine.stop() / machine.delete() / machine.state()
• Use it as a context manager to auto-delete() on exit.
• Errors are typed: SmolError (with .code), ExecutionError, NotSupportedError, InvalidConfigError.

ExecResult has .exit_code, .stdout, .stderr, .success, .output, and .assert_success().

Install / build from source

The cloud path is pure Python. The local path needs the native extension, which links libkrun from the sibling smolvm repo (three levels up).

python -m venv .venv && . .venv/bin/activate
pip install maturin
# Build + install the native extension (points at the repo's bundled libkrun):
LIBKRUN_BUNDLE=../../../lib maturin develop

To boot local microVMs the engine needs a code-signed boot helper carrying the macOS com.apple.security.hypervisor entitlement (the Python process itself does not). Point it at one (and the libkrun dir):

SMOLVM_BOOT_BINARY=../../../target/release/smolvm \
SMOLVM_LIB_DIR=../../../lib \
python your_script.py

On Linux the host needs /dev/kvm.

Tests

python tests/test_unit.py        # error parsing + path encoding (no VM/network)
python tests/test_cloud_mock.py  # cloud transport vs a mock /v1 (no VM/network)
python tests/test_async_mock.py  # AsyncMachine vs a mock /v1 (concurrency, no VM/network)
# Local VM boot (needs the native build + the env above):
SMOLVM_BOOT_BINARY=… SMOLVM_LIB_DIR=… .venv/bin/python tests/test_local_e2e.py

License

Apache-2.0