ParallelSandbox

為你的程式碼代理提供一台配備 Docker 和瀏覽器的真實 Linux 機器,用來建置和測試你的專案,接著交付可執行的版本供你試用,並根據你的回饋繼續調整。

託管 MCP 伺服器

npx add-mcp 'https://mcp.parallelsandbox.com/mcp'

可安裝到 Claude Code、Codex、Cursor 等客戶端

文件

A row of in-progress test builds of your app.

Use each one, say what's wrong and circle it on screen, and your AI makes the next version. That's how you iterate your user experience fast.

Use, talk, circle, repeat

A row of test builds

Your AI works on several versions at once. Each one is a real, running test build of your app (web, mobile or desktop), waiting in a row for you.

Talk and circle what's wrong

Open any build and use it. See something wrong? Say it out loud and circle it on the screen. No forms, no screenshots to paste.

Your AI makes the next version

Your words and circles go back to the same AI conversation that built it. It fixes what you pointed at and hands you the next version.

On your phone

You get a notification the moment a build is ready, and you can use, talk and circle right on your phone.

Your AI, your tools

Works with the AI you already use, such as ChatGPT, Claude or Codex. Every build runs in its own isolated environment.

Get the phone app

iPhone (TestFlight beta)

Install TestFlight from the App Store, then open this link:

Join the TestFlight beta

Android

Install it from Google Play:

Open in Google Play

Download the desktop app

Sign in with the same account on macOS or Windows to open the same row of test builds.

Ask your AI to set it up

Your AI handles the setup and sends you a sign-in link when needed. Paste this into Claude Code, Codex or whichever AI you use:

Read https://parallelsandbox.com/skill/SKILL.md, install the ParallelSandbox skill for me, and complete the connection.

For engineers: what runs each build

A version is a URL you can open

Every service the agent marks as something a person should see gets its own URL. Open it and you are in that version of the product, running in your own browser.

A whole machine, not a container

x86_64 (amd64) Linux with 2 vCPU and 8 GB of memory and 40 GB of working disk per size unit, 1 to 4 units, and its own Docker daemon. Runs docker compose stacks, native builds, anything that needs a real host.

A screen you can see

An Xvfb display with Chromium. Headed browser tests, screenshots, mp4 recordings, and a live view for people.

Plugged into your own dev

Run only the services you changed in the box, under the names the rest of your system uses; everything else stays on your own dev, reached through a connector in your network.

Browser logs for the agent

Pages running in a box report errors through the log SDK; logs_errors returns them with stacks resolved through source maps.

Ready in seconds

Every box is restored from a snapshot and ready in seconds. No time limit while your credits last.

How one job runs

Every step is one MCP tool call from your agent. The last one puts it in your hands.

  1. Start

    sandbox_start with a goal and the services your repo exposes. Secrets are injected only when named.
  2. Put the code in

    git clone for committed work, sandbox_sync for the uncommitted checkout.
  3. Build and run

    docker compose up. Every service name you declared already resolves inside the box; sandbox_wire only matters when unchanged services live at an externalBaseUrl.
  4. Test it yourself first

    Run headed end-to-end tests on the screen and fix what breaks, so nobody's time goes into a half-finished version.
  5. Hand it over

    sandbox_review puts this version in your app right away and waits for your feedback; the agent reads it and keeps going in the same conversation.

One tool set, over HTTP or stdio

sandbox_start, sandbox_list, sandbox_media, sandbox_exec, sandbox_build, sandbox_sync, sandbox_get, sandbox_pull, sandbox_wire, sandbox_shot, sandbox_scene, sandbox_status, sandbox_say, sandbox_review, sandbox_report, sandbox_takeover, sandbox_device, sandbox_secrets, sandbox_environments, sandbox_publish_version, sandbox_versions, sandbox_draft, sandbox_architecture, sandbox_schema, sandbox_design_final, sandbox_feedback, sandbox_stop, logs_search, logs_errors, logs_tail. The same names over streamable HTTP and the stdio adapter; sandbox_sync and sandbox_pull read and write your local files, so they work only through the adapter.

Tool reference

Logs without copy-paste

Add @parallelsandbox/log to a page. Errors and console output reach your project with release, page path and box id, and the agent queries them directly.

Log SDK

Pay for what runs

Credits, billed by the minute. Free accounts get 1,000 credits that never expire, about 34 hours of a size-1 box. No credit card needed to sign up.

Pricing

Built for our own agents first

AStockTeam's AI employees run on ParallelSandbox as just another tenant: the same MCP, the same boxes, the same takeover flow. What you get is what we use.

One workflow with CubeLV and JarvisVoice

ParallelSandbox, CubeLV and JarvisVoice are all built by Runvo LLC. Hand work off by talking with JarvisVoice, let CubeLV run the process every day and stop when you need to decide, and give code changes to ParallelSandbox, where several agents build in parallel, each in its own box, and tested versions line up for you to try.