Sanbox - Resumable MicroVM Sandboxes for AI Agents

TL;DR

TL;DR: Sanbox runs AI agents in isolated, resumable MicroVM sandboxes with persistent filesystems and live run events. The CLI integrates with Claude Code, Cursor, Codex, and CI — with a self-hosting option.

What Is Sanbox?

Sanbox is a platform for running AI agents in isolated, resumable sandboxes. Each sandbox is backed by a MicroVM that provides strong isolation from the host system, a persistent filesystem that survives agent restarts, and live run events so you can observe what the agent is doing in real time.

The project was shared on Hacker News as “Show HN: Sanbox, batteries included sandboxes for AI agents” (HN thread).

Key capabilities:

• MicroVM isolation — each agent runs in its own lightweight VM, isolated from the host
• Resumable sandboxes — a sandbox can be paused and resumed; the agent picks up where it left off
• Persistent filesystem — files persist across runs without needing a shared volume
• Live run events — stream agent activity in real time (file changes, tool calls, LLM responses)
• CLI-first — works with Claude Code, Cursor, Codex, GitHub Actions, or a plain terminal
• Self-hosting option — deploy Sanbox on your own infrastructure

Setup Workflow

Prerequisites

• macOS, Linux, or a VPS with Docker (for self-hosted)
• Node.js 18+ (for the CLI)

Step 1: Install the CLI

npm install -g @sanbox/cli

Step 2: Authenticate

sanbox login

This opens a browser for OAuth. Cloud-hosted sandboxes require a free account.

Step 3: Create a sandbox and run an agent

sanbox run --agent claude-code "Summarize the files in this project"

Supported --agent values include claude-code, cursor, codex, and terminal (raw shell).

Step 4: Stream live events

sanbox events <sandbox-id>

Returns a real-time feed of file changes, tool invocations, and LLM responses as the agent works.

Step 5: Self-host (optional)

The self-hosted option requires a Linux machine with KVM support. Docker is not used — Sanbox manages MicroVMs directly via firecracker.

sanbox server start --self-hosted

Point the CLI at your self-hosted endpoint:

sanbox config set endpoint https://your-server:3000

How It Works

OpenCode SDK

Sanbox uses the OpenCode SDK as its agent harness. OpenCode is an open-source framework that connects coding agents to tool chains. Sanbox wraps OpenCode sandboxes in MicroVMs, adding isolation and resumability.

OpenCode is the default harness, but you can bring your own:

sanbox run --harness custom -- your-harness-command

Template System

Sanbox supports reusable sandbox templates — pre-configured environments with specific tools, environment variables, and startup scripts.

# Save current sandbox as a template
sanbox template save my-template

# Launch from template
sanbox run --template my-template "Analyze this dataset"

Resumability

When an agent task is interrupted (Ctrl+C, timeout, or crash), the sandbox state is preserved. Running the same sandbox ID resumes the agent with full memory of what it already did.

sanbox resume <sandbox-id>

Practical Evaluation Checklist

• Can you create a sandbox and run a Claude Code agent in under 2 minutes?
• Does sanbox events show live output as the agent works?
• Does sanbox resume restore the agent state after an interruption?
• Is the self-hosted server able to start on a fresh Ubuntu 22.04 machine with KVM?
• Are the persistent filesystem writes visible after a sandbox restart?

Security Notes

MicroVM isolation means each sandbox has its own kernel, network stack, and filesystem — a compromised agent cannot access the host system. This is significantly stronger than container-based isolation (Docker), which shares the host kernel.

Network access is controlled by per-sandbox ACLs. By default, outbound network is allowed; you can restrict it to specific domains.

For self-hosted deployments, ensure KVM (/dev/kvm) is accessible to the Sanbox process and that the firewall blocks external access to port 3000 unless behind a reverse proxy with auth.

FAQ

Q: What is the difference between Sanbox and running Claude Code or Cursor locally? A: Local agents run directly on your machine with full access to your filesystem and network. Sanbox sandboxes are isolated MicroVMs — an agent gone rogue cannot reach your home directory or internal services. Useful for testing untrusted agent code or running many agents in parallel without interference.

Q: Is there a free tier? A: The website mentions 500 free sandboxes per month on the cloud-hosted plan, with self-hosting available at no cost on your own hardware.

Q: Does Sanbox work with other agents besides Claude Code and Cursor? A: Yes. The CLI also supports OpenAI Codex and a plain terminal mode. Custom harnesses can be plugged in via the --harness flag.

Q: What does self-hosting require? A: A Linux machine with KVM enabled (nested virt not required), at least 4 GB RAM per concurrent sandbox, and Ubuntu 22.04 or Debian 12. macOS is supported for the CLI only.

Q: How is resumability implemented? A: The MicroVM state (memory snapshot + filesystem diff) is saved to disk between runs. On resume, the VM is restored from the snapshot and the filesystem diff is replayed, giving the agent a consistent view of its working environment.

Source and Accuracy Notes

• Project page: sanbox.cloud
• HN launch thread: news.ycombinator.com/item?id=48879908
• Source last checked: 2026-07-20