A cloud machine
for your AI agent.
Isolated Linux sandboxes your agent controls over HTTP. Start with one CLI command, verify by one text, then run commands, install packages, and read or write files. $100 of compute free.
For agents: fresh setup starts with this command. API keys are created by the flow, and the verified phone is inferred from the SMS sender rather than typed into the terminal.
# First, verify by one human text # No API keys or phone number needed up front $ npx -y @sendblue/cli@latest sandbox init --no-wait → text "SB SETUP 123456" to +1... $ npx -y @sendblue/cli@latest setup --check ✓ credentials saved $ npx -y @sendblue/cli@latest show-keys → API Key / API Secret # Then spin up a sandbox $ npx -y @sendblue/cli@latest sandbox create ✓ { "id": "sbx_a9c7...", "status": "running" } # Run a command in it $ curl -X POST ".../sandboxes/sbx_a9c7.../exec" \ -d '{"command": "git clone … && npm test"}' ✓ exit_code: 0
Sign up. Create.
Hand it to your agent.
From zero to a running sandbox your agent can drive, in three steps.
Bootstrap with one text
Run npx -y @sendblue/cli@latest sandbox init. Fresh setup has no API keys or phone number entered up front: the CLI shows a Sendblue number and setup code, and the human texts that exact code from the phone they want to verify. That unlocks $100 of free compute and creates the first sandbox.
Create a sandbox
Use sendblue sandbox create or call POST /v3/sandboxes with your API key headers. You get a sandbox id back in seconds, ready to run commands.
Connect your agent
Call GET /v3/sandboxes and paste the connect.prompt into your agent’s instructions. Now it can create, run, and read on its own.
Read results out, destroy when done
Pull files back over the API before the sandbox sleeps, then DELETE it — the unused portion of the awake window is refunded.
# 1 · Create $ curl -X POST .../v3/sandboxes ✓ sbx_a9c7 running # 2 · Work (shell state persists while awake) $ ... /exec -d '{"command":"cd /root && git clone …"}' $ ... /exec -d '{"command":"npm test"}' ✓ exit_code: 0 # 3 · Read results out before it sleeps $ curl ".../files?path=/root/report.json" → saved locally # 4 · Destroy (refunds unused time) $ curl -X DELETE .../sandboxes/sbx_a9c7 ✓ destroyed
A real machine,
one API call away.
Everything your agent needs to run code safely in the cloud — no infrastructure to manage.
Run any command
Full shell access on Ubuntu Linux (x86_64) as root, with Node 22, Bun, and git preinstalled. apt installs anything else — Python included, in about ten seconds.
Read & write files
Push files into the sandbox and pull results back out over the API. Perfect for feeding an agent a script and reading whatever it produces.
Fully isolated
Every sandbox is its own container, walled off from every other account. Safe for running untrusted or AI-generated code.
Sleeps when idle
Sandboxes sleep automatically after 10 minutes and wake on the next command or file operation. You’re only billed for time a sandbox is actually awake.
Metered by the second
$0.33 per awake-hour, and destroying a sandbox refunds the unused window. A typical short session costs a few cents.
Just HTTP
No SDK required. Six REST endpoints, standard API-key headers, JSON in and out. Works from any language or agent framework.
Copy one prompt,
hand it to any agent.
Call GET /v3/sandboxes and the API hands back a ready-made connect prompt. Paste it into Claude, GPT, or your own agent along with your keys — and it drives the sandbox on its own.
- Create and destroy sandboxes on demand
- Run shell commands and read the output
- Write files in, read results back out
- Check remaining compute balance any time
Bring your own model — sandboxes are the compute, your agent is the brain.
# Get the ready-made agent prompt $ curl ".../v3/sandboxes" -H "sb-api-..." { "connect": { "prompt": "You have access to a Sendblue sandbox: an isolated Linux container you control over HTTP..." } } # Paste that into your agent. Done. > "Clone my repo and run the tests" ✓ agent runs it in the sandbox
What people build
when agents get compute.
Anywhere an agent needs to actually run code, not just talk about it.
Coding agents
Let an AI agent clone a repo, install dependencies, run the test suite, and iterate — in a clean machine that can’t touch your infra.
Run untrusted code
Execute AI-generated or user-submitted code safely. Each run is isolated in its own container and torn down when you’re done.
Data & scripts
Spin up a box to crunch a dataset, run a one-off script, or process a file — then destroy it and pay only for the seconds it ran.
Tool use for LLMs
Give your LLM a real code interpreter: it writes a script, runs it in the sandbox, reads the result, and keeps going.
CI-style checks
Fire off a build or lint in a fresh environment on demand, without maintaining your own runners or worrying about state carryover.
Prototyping & demos
Test an idea in a real Linux environment in seconds — no VM to provision, no Docker to configure, no cleanup afterward.
$100 free to start.
Pay only while awake.
Verify your phone number and get $100 of sandbox compute free — no credit card required. After that it’s $0.33 per awake-hour, with unused time refunded on destroy.
- Sleeping sandboxes cost nothing
- Destroy refunds the unused awake window
- A typical short session costs a few cents
$ curl ".../v3/sandboxes" -H "sb-api-..." balance $99.94 remaining sandboxes 1 active · sleeping · awake time bills at $0.33/hour · sleeping costs $0.00
Frequently asked questions
If something’s not covered here, the docs usually have it.
What is a Sendblue sandbox?
A sandbox is an isolated cloud Linux machine (Ubuntu, x86_64, running as root) that your AI agent controls over a simple HTTP API. It can run shell commands, install packages with apt, clone repos, and read or write files — fully walled off from every other account. Sandboxes boot in seconds and sleep automatically when idle, so you only pay for the time you actually use.
How do I get $100 of free sandbox compute?
Run sendblue sandbox init, then text the shown setup code from the standard mobile phone you want to verify. The phone number is inferred from the SMS sender, not typed into setup. Once verified, your account gets $100 of sandbox compute free. Each phone number can claim the free tier once.
How is sandbox compute billed?
Compute is metered at $0.33 per awake-hour. Sandboxes sleep automatically after 10 minutes of inactivity, and sleeping sandboxes cost nothing. Destroying a sandbox refunds the unused portion of its final awake window, so a typical short session costs just a few cents.
How does my agent connect to a sandbox?
Call GET /v3/sandboxes and copy the connect.prompt field into your agent’s instructions along with your API keys. Any agent that can make HTTP requests — Claude, GPT, or your own — can then create sandboxes, run commands, and read results on its own. You bring the model; sandboxes are the compute.
Do sandbox files persist?
Files persist while a sandbox is awake — any command or file operation on that sandbox resets the 10-minute idle timer (listing sandboxes does not). When it sleeps, the filesystem resets, so treat each working session as a unit of work and read your results out through the files API (or push them somewhere) before going idle.
Give your agent
a machine of its own.
Create your first sandbox in seconds — POST /v3/sandboxes is all it takes.