I've been obsessed with the rise of "computer-use agents" lately—tools like Claude Code that don't just autocomplete but actually interact with the OS.

The problem is infra: giving an agent access to your host is a nightmare, and spinning up full cloud VMs for every task is slow and expensive. I spent last night building Boxed to solve this. It's an open-source (MIT) execution engine that gives agents a sovereign, ephemeral OS layer—fully under your control.

Why build it? I wanted something that handles the compute heavy-lifting (sub-second boot, native artifact streaming, persistent REPL sessions) so I could focus on building the agents themselves. No more wrestling with sandboxes that feel bolted-on.

Usage: It's simple to integrate. You run the Boxed control plane locally or in a container, and use the SDKs to spawn isolated sessions.

Example (Python):

python

from boxed_sdk import Boxed client = Boxed(base_url="http://localhost:8080", api_key="your-secret") session = client.create_session(template="python:3.10-slim") # Run code with ephemeral OS access & artifact capture result = session.run("import pandas; print('Processing data...')") print(result.stdout) session.close()

typescript import { Boxed } from '@boxed/sdk';

const client = new Boxed({ baseUrl: 'http://localhost:8080', apiKey: 'secret' }); const session = await client.createSession({ template: 'python:3.10-slim' });

// Runs in a secure, ephemeral environment const { stdout } = await session.run("print('hello from HN')"); console.log(stdout);

await session.close();

Tech: Control Plane: Built in Go for a high-performance, concurrent REST API. Agent: A lightweight Rust binary injected into every sandbox to manage lifecycles and stream results via WebSockets. Isolation: Docker-based for now (with Firecracker MicroVMs on the roadmap for Phase 2).

It's early days, but it’s running smooth in my own agent workflows. I’d love your feedback on the architecture

Repo: https://github.com/akshayaggarwal99/boxed

Demo and quickstart are in the README.