New:Upstash has a remote MCP server
Upstash

Cloud computers for AI agents

Upstash Box is a persistent cloud computer with coding agents built in. Run Claude Code, Codex or OpenCode in an isolated box, give it a browser and a repository, and get results back through an API. The workspace stays between tasks, you pay only for the CPU your agent actually uses, and your own coding agent can work in a box through the Upstash MCP server.

Why Upstash Box

Everything an agent needs to finish real work, without building the infrastructure around it yourself.

Agents built in

Claude Code, Codex, OpenCode or Cursor run inside the box. Send a prompt and get the result back with its cost, as text, a stream, or typed output.

Persistent workspace

The repository, installed dependencies and files stay between tasks. A user comes back tomorrow and picks up where they left off.

Pay for active CPU

Agents spend most of their time waiting on the model. Box bills only the CPU your code uses, memory is free, and an idle box pauses.

Headless browser

Chromium in the same box: open the app the agent built, take screenshots, act on the page and extract typed data, with live view and recordings.

Git to pull request

Clone, branch, commit, push and open a pull request from the SDK, so a finished task arrives as a PR instead of a diff in a container.

Schedules and always-on

Run a command or an agent prompt on a cron schedule inside the box, or keep a box on all the time for a flat monthly price.

Isolated by default

Each box has its own filesystem, processes and network, with no access to private networks. Allowlist outbound traffic and keep API keys off the box.

Snapshots and fan-out

Snapshot a ready environment once, then start new boxes from it for each user, task, or parallel attempt.

Give your agent a computer

Build with the @upstash/box SDK for TypeScript and Python, the Box CLI, or the Upstash MCP server.

import { Box, Agent } from "@upstash/box"

const box = await Box.create({
  runtime: "node",
  agent: {
    harness: Agent.ClaudeCode,
    model: "anthropic/claude-opus-5-5",
    apiKey: process.env.ANTHROPIC_API_KEY,
  },
})

const run = await box.agent.run({
  prompt: "Set up a Next.js project with Tailwind",
})

console.log(run.result)

Use Box from your coding agent

Connect Claude Code, Codex, Cursor or OpenCode to the Upstash MCP server and your agent can create boxes, run commands, open pull requests, publish preview URLs and drive a browser, all on a remote machine instead of your laptop.

Run in your terminal, then type /mcp and choose Authenticate

claude mcp add --scope user --transport http upstash https://mcp.upstash.com/mcp

Server URL: https://mcp.upstash.com/mcp. Sign in with OAuth on the first tool call, or pass an API key for CI. MCP setup guide

What can you build with Box?

Agent workloads teams run on Box today, with a guide for each.

Background coding agents

Hand off a task, close the laptop, and review a pull request later. The agent clones, fixes, tests and opens the PR in its own box.

Parallel coding agents

One box per task, all started from the same snapshot, so agents never share a checkout or break each other's builds.

Per-user agent workspaces

Give every user of your product their own box that keeps their project between sessions and pauses when they leave.

Personal agents, always on

Host OpenClaw or Hermes on a keep-alive box that stays reachable from Telegram, WhatsApp or Slack while your laptop is off.

Scheduled browser testing

A box with a browser and a cron schedule wakes up, runs Playwright against your app, saves the results and pauses again.

Remote work from your coding agent

Connect Claude Code, Cursor or Codex to the Upstash MCP server and the agent builds, tests and opens pull requests on a box instead of your laptop.

What is an AI agent sandbox?

A computer the agent can use, without access to yours.

An AI agent sandbox is an isolated environment where an agent runs the code it writes. Coding agents like Claude Code and Codex install packages, run builds and execute commands with real permissions. On a laptop or a shared server, one wrong command or one malicious package can reach your files, your credentials and your network. In a sandbox, the worst it can break is a disposable machine.

Most sandboxes stop there: they give you an isolated machine and leave the rest to you. You install the agent, stream and parse its output, keep its files between sessions, run git, add a browser, and track what each run costs. That integration work is where the time goes.

Upstash Box ships it. Each box is an isolated container with the coding agent already inside, a workspace that persists, git with pull requests, a headless browser and cron schedules, all driven from the SDK. Boxes pause when idle and bill active CPU only, so an agent that spends most of its time waiting on the model costs little to keep around.

Frequently asked questions

Upstash Box is a persistent cloud computer for AI agents, with coding agents built in. Each box is an isolated Linux container with a shell, a filesystem that persists, git, and a headless browser. You create one with the TypeScript or Python SDK, give the agent inside a task, and get the result back through an API.

An AI agent sandbox is an isolated environment where an agent runs the code it writes. It has its own filesystem, processes and network, so a wrong command or a malicious package cannot reach your machine, your data or other users. Some sandboxes only run code you send them; Upstash Box also runs the agent itself inside the sandbox.

Claude Code, Codex, OpenCode and Cursor are built in. You choose the agent and the model when you create a box and run it with box.agent.run(). You can also bring your own harness, or skip the built-in agent and drive the box's shell, file and git APIs from your own agent loop.

Yes. Create a box with Claude Code as its agent and send it a prompt. It works in the box's own workspace, can clone your repository, run tests and open a pull request, and every run reports what it cost. It keeps working after you close your laptop.

Yes. Add the Upstash MCP server at https://mcp.upstash.com/mcp to Claude Code, Cursor, Codex, OpenCode or any MCP client and sign in with OAuth. Your agent can then create and manage boxes, run commands, use git and open pull requests, publish preview URLs for ports, drive a headless browser and set up schedules, so the work runs on a remote box instead of your laptop.

Yes. A standard box pauses automatically when it is idle. Its storage is preserved, including the cloned repository, installed dependencies and files, but running processes stop until the next request wakes the box. Snapshots save a box's disk so new boxes can start from it.

Yes. A keep-alive box never pauses, so a personal agent, a bot or a server always has a warm machine. Keep-alive boxes have a fixed monthly price by size: $8 for Small, $16 for Medium and $32 for Large. An init command can start your process every time the box boots.

The free plan includes up to 10 concurrent boxes, 5 CPU hours per month and $1 of LLM usage per month. Pay as you go bills active CPU only: $0.10, $0.20 or $0.40 per active CPU hour for Small, Medium and Large boxes. Memory is not billed, and an idle box costs nothing for CPU. Storage is $0.10 per GB per month, including snapshots. Keep-alive boxes are $8 to $32 per month with no usage charges. See the pricing page for details.

Every box runs in its own container with a dedicated filesystem, process tree and network stack. Boxes cannot reach or observe each other, and they cannot reach private networks or cloud metadata services. A network policy can limit outbound traffic to an allowlist, and Attach Headers inject secret HTTP headers on the host so API keys never enter the box.

Yes. Bring your own key for any LLM provider on every plan. Without one, boxes use Upstash-provided model usage, capped at $1 per month on the free plan and $100 per month on pay as you go.

Boxes currently run on AWS us-east-1, with more regions planned. Boxes come in three sizes, from 2 vCPU and 4 GB of memory up to 8 vCPU and 16 GB, with Node.js, Python, Go, Ruby and Rust runtimes. Custom Docker images are planned.

Those give you an isolated machine and leave the agent loop to you. Box runs the coding agent inside the sandbox, with git-to-pull-request, a browser and schedules built in, and bills active CPU with memory free. E2B and Vercel Sandbox use microVMs with their own kernel, a stronger boundary than Box's container. Daytona offers GPUs, and both E2B and Daytona offer EU regions. Our comparison of 15 sandbox providers covers the details.

Run your first agent in minutes

Start on the free tier with 10 concurrent boxes, 5 CPU hours a month and $1 of model usage. No credit card required.

Create BoxView Pricing