Bot Perch

Use casesCoding agent

Give a coding agent its own machine

A coding agent reads a repository, edits files and runs tests. In a sandbox, a mistake cannot touch your laptop, and the work carries on when the laptop is closed. Several providers ship Claude Code and Codex already installed.

Short answer: The cheapest machine that runs it as described is Agent 37 at $6.76 a month, then Runtime ($13.49) and Lizard ($19.00). 12 of 16 providers run it without a catch. Machine cost only, before any AI model, at list prices checked 29 October 2026.

Setup checked 29 October 2026Docs linked in the textPriced from list rates

What it is

“I want Claude Code or Codex to work on my code in the cloud, not on my laptop.”

Tools like Claude Code and Codex run in a terminal: you give them a task, they read the code, change it and run commands to check their work. Running them in a cloud sandbox gives each one a disposable Linux machine with your repository on it.

Unlike a personal assistant, a coding agent does not need to stay on. It runs while you work and can stop at night, as long as the repository, its caches and the agent's login survive.

What it needs from a machine

From its own documentation, each line linked to the page it comes from

Machine size

Claude Code needs 4 GB of memory or more. Anthropic's own cloud sessions allow up to about 4 vCPU, 16 GB and 30 GB of disk. 4 vCPU / 8 GiB / 40 GiB covers the agent, tests and a few containers.

code.claude.com

Does it stay on?

No. It runs during working hours; eight hours on 20 days is 160 hours a month.

battleships.dev, built by the Boat team

What must survive a restart

The repository and uncommitted work, the agent's login (~/.claude for Claude Code), and caches and Docker images if you want a fast start.

code.claude.com

Browser, Docker, GPU

Docker helps when tests need databases or services. A browser matters only for checking web interfaces. No GPU.

battleships.dev, built by the Boat team

How busy it is

Mostly waiting on the model: about 4% CPU on average in Battleships' software-factory workload, with short peaks during builds.

battleships.dev, built by the Boat team

Setting it up on each provider

Easiest first. “Without a terminal” means someone who does not type commands can get it running from the provider's website.

ProviderHowWithout a terminal?What it involves
Agent 37One clickYesClaude Code and Codex are on the one-click Personal plans, on your own Anthropic or OpenAI account.Agent 37 Personal ↗
BoatPreinstalledNoSeven coding agents on every sandbox, including Claude Code and Codex; boat prompt runs one for you.Boat: harnesses and models ↗
boxdPreinstalledNoEvery machine ships with Claude Code, Codex and OpenCode.boxd: coding agents ↗
InstaBoxPreinstalledPartlyClaude Code and Codex come installed, with a desktop you can open in the browser.InstaBox: limits and images ↗
LizardPreinstalledPartlyChoose the claude or codex template when creating a sandbox.Lizard: sandboxes ↗
SandboxerPreinstalledNoClaude Code, Codex, OpenCode and others are installed on every machine.Sandboxer: machines ↗
Cloudflare SandboxOfficial guide or templateNoAn official guide: Claude Code edits a GitHub repository in a sandbox and returns a diff.Cloudflare: Claude Code ↗
DaytonaOfficial guide or templateNoAn official guide for running Claude Code in a sandbox from the Daytona command line.Daytona: Claude Code ↗
E2BOfficial guide or templateNoAn official guide for running Claude Code, Codex and Amp in E2B sandboxes from code.E2B: coding agents ↗
TensorlakeOfficial guide or templateNoThe first-party route is the `tensorlake-opencode` plugin: OpenCode keeps running on your laptop and its bash, read, write, edit and patch calls run in a sandbox that survives restarts and can mount a volume or a hosted Git repository. Claude Code and Codex are not in the images, but they work over SSH or VS Code Remote-SSH, and Tensorlake has self-hosted guides for Claude Managed Agents, Cursor Cloud Agents and Devin Outposts.Run OpenCode with Tensorlake Sandboxes ↗
Vercel SandboxOfficial guide or templateNoEach Claude Code session gets its own sandbox, through Vercel's agent harness.Vercel: Claude Code ↗

Read from each provider's documentation on 29 October 2026. A provider not listed as one click can still be the right home; it means someone has to follow its instructions.

Where to start

Three starting points for different people, each with its catch

One machine per developer

Boat

Claude Code and Codex preinstalled on a full VM with Docker. The 4 vCPU / 8 GB size for 160 hours is about $6 of usage, inside the $20 plan.

Catch: Auto-stop is on by default, which suits this job, and a stopped disk is kept for 14 days.

docs.boat.dev ↗

If you want memory billed on use

boxd

Three coding agents preinstalled, memory billed on what is actually in use, and a 100 GiB disk billed only for what is written.

Catch: CPU is billed on the full size while running, and the 4 vCPU size comes with 16 GiB.

docs.boxd.sh ↗

One fresh sandbox per task, from code

E2B

Built for starting sandboxes from your own program, with an official coding-agent guide and pausing that keeps memory and files at no charge.

Catch: Disk is 10 GiB on the free plan and 20 GiB on Pro, and sessions longer than an hour need the $150 Pro plan.

e2b.dev ↗

Beyond these providers. Anthropic runs Claude Code cloud sessions on its own machines for Claude subscribers, and OpenAI runs Codex Cloud for ChatGPT plans. Neither publishes a separate machine price on its docs page.

Compare these three head to head: features and cost side by side.

What it would cost

One 4 vCPU / 8 GiB / 40 GiB machine for Claude Code or Codex, eight hours a day on 20 working days, stopped otherwise with the repository kept. The CPU works about 4% of the time on average. Change any number; open a row for the arithmetic.

Start from

The machine
How it runs
How busy it is

Only providers that bill for CPU actually used look at this. An agent waiting for messages is near 0%; a build or a test run is near 100%.

For providers that put an idle machine to sleep and wake it on a request: the share of its hours it has to be awake. An agent that only answers webhooks can sleep; one that holds a chat connection open cannot.

The AI model (optional)

For most agents the model costs more than the machine. Each message re-sends the agent's instructions and history, so tens of thousands of tokens is normal. Prices are the median of the providers that keep no data, from TokenWatch, read 2026-10-29.

    List prices only, excluding model tokens, egressData the machine sends out to the internet, such as downloads to a user or uploads to another service. Some providers charge per GB; most agents send very little., taxes and anything a provider does not publish. How the numbers are made.

    The features that decide it

    Each page compares every provider, with a source for every row

    1. 01

      Isolation: what kind of computer you get — Every provider says “sandbox”. Underneath, it is a container, a gVisor container, a microVM or a full virtual machine, and that decides what your agent can install, run and break.

    2. 02

      Snapshots, forks and templates — Four jobs hide under “snapshot”: a restore point on one machine, a reusable image for new machines, a copy of a running machine, and a backup that survives losing the host. Ask which one you are getting.

    3. 03

      Lifecycle: stop, pause and sleep — Providers use seven words for four behaviours. The only question that matters is what is still running when the machine comes back.

    4. 04

      Built-in agents — Some providers ship the machine with coding agents already installed and a way to prompt them. Others ship a computer and expect you to bring the agent.

    5. 05

      Sizes, regions and hardware — A size is a bundle of CPU, memory and disk, and the bundle decides what you can run. A region decides how far away it is and whose laws it sits under. Neither is as flexible as the word “cloud” suggests.

    Sources