update
Jun 4, 2026
By Teun
Boxes.dev puts Claude Code and Codex in cloud VMs
boxes.dev launched a service that runs each Codex or Claude Code thread on its own cloud Linux VM, called a fork. The company said the setup keeps agents running in parallel, preserves services and database state, and can be controlled from desktop or mobile.
boxes.dev has launched a cloud workbench for AI coding agents that gives each Codex or Claude Code thread its own Linux virtual machine. The company said the system is designed to move coding work off a developer’s laptop and into the cloud, so agents can keep running while the user is away.
According to the product site, boxes.dev treats each unit of work as a “fork,” which is a full computer rather than a simple git worktree. Each fork includes its own filesystem, local services, port forwarding to localhost, and preinstalled development environment.
⚡ New to this?
This matters because a lot of AI coding tools still assume one developer, one machine, and one active session. Boxes.dev is trying to change that by giving each agent its own cloud computer, so “parallel agents” means multiple AI workers can run at once without sharing the same setup.
A virtual machine, or VM, is a software-based computer running in the cloud. For non-specialists, the key point is that it can keep code, services, and database state separate while still behaving like a real machine a developer can inspect and control.
🦞 OpenClaw angle
If you run self-hosted agents, separate them by task instead of letting multiple jobs share one workspace or container. Use isolated VMs or similar per-job environments when agents need to edit code, run services, or touch databases, because that avoids cross-thread interference.
If your automation needs approvals or long-running work, add a control path that works when the main laptop is closed: web, mobile, or API review steps, plus persistent logs and diffs. For team rollouts, plan for RBAC, audit logs, spend limits, and network allowlists from the start rather than bolting them on later.
The company said that setup is meant for parallel work. Developers can spin up multiple agents at once, with each one isolated on its own machine, instead of sharing one environment and stepping on each other’s files, ports, or database state.
The service also keeps working after a laptop closes. boxes.dev said users can start a task, leave it overnight, and later review the results from desktop, web, or mobile, including iPhone and Android. The company said users can read threads, inspect diffs, answer questions, approve requests, and even open a terminal from a phone.
boxes.dev said it supports Codex and Claude Code using the subscriptions users already have. In the setup flow shown on the site, the system scans a local codebase, generates a setup plan, and then recreates the developer environment in the cloud before forking it for new tasks.
The company described three basic steps: move a project into a cloud devbox, run each task on its own machine, and then review or commit the changes from anywhere. It said local port forwarding follows the selected fork, which lets developers test against localhost while working in the cloud.
The product page also lists scheduled runs and integrations. boxes.dev said tasks can be triggered from Slack or Linear, from a pull request, or on a schedule, such as a weekly dependency audit that opens a PR when it finishes.
The company said the service uses a usage-based model tied to “box-hours,” which it defined as one 4 vCPU / 8 GiB devbox awake for one hour. It offers a free trial with 10 box-hours, plus paid Starter, Pro, and custom Teams plans.
For larger organizations, boxes.dev said the Teams tier adds SSO, SCIM, role-based access control, audit logs, spend controls, network allowlists, dedicated resource pools, BYO cloud or VPC support, and security review options. The company said it built the product because other tools only cover parts of the agent workflow, while boxes.dev is meant to give each agent a full computer in the cloud.
The service is now available through a desktop app, CLI, and mobile apps, and the company is pitching it as a way to run coding agents in parallel without keeping a laptop open.
