vpod 0.2.0

Lightweight, secure sandboxes for untrusted processes.

vpod Python SDK

A lightweight, portable sandbox that gives an untrusted process an instant Linux environment. It uses the RISC‑V architecture and runs entirely inside WebAssembly.

• Fast startup — Boot in under a second.
• Portable — Runs anywhere without any setup required.
• Isolated — All execution state stays inside the WASM sandbox.

Installation

pip install vpod

Usage

Shell commands (stateless)

Each call gets a fresh VM — no shared state:

from vpod import Sandbox

sbx = Sandbox.create()
result = sbx.commands.run("echo hello")
print(result.stdout)    # hello
print(result.exit_code) # 0

Persistent session

All calls share the same running VM:

from vpod import Sandbox

with Sandbox.create() as sbx:
    sbx.commands.run("export Foo=Bar")
    sbx.commands.run("touch /tmp/data.csv")

    result = sbx.commands.run("echo $Foo")
    print(result.stdout)  # Bar

Python REPL

Run Python code with persistent state across calls:

from vpod import Sandbox

with Sandbox.create() as sbx:
    sbx.code.run("import requests")
    sbx.code.run("data = [1, 2, 3]")
    result = sbx.code.run("print(sum(data))")
    print(result.text)  # 6

Variables and imports persist for the lifetime of the session.

Snapshots

The first call to Sandbox.create() downloads the VM snapshot (~50MB) and caches it locally at ~/.local/share/vpod/snapshots/. Subsequent calls use the cache instantly.

To pre-download (e.g. in a Dockerfile or CI setup):

from vpod import snapshots

for s in snapshots.fetch_registry():
    print(s["name"], s["tag"])

path = snapshots.pull("alpine:latest")

How it works

A vpod runs a RISC‑V virtual machine compiled to WebAssembly. The core implements the RV64GC specification:

• G (General-purpose): I/M/A/F/D extensions for integer, multiply/divide, atomics, and floating-point
• C (Compressed): 30% smaller code size, improving memory efficiency

The WASM component communicates with the host through WASI 0.2, providing controlled access to networking and I/O while keeping all execution state isolated inside the sandbox.

Limitations

• Emulation overhead — No hardware acceleration in WASM. CPU-intensive workloads run slower than native.
• No GPU access — CUDA, Metal, and hardware ML accelerators are not yet available.