Sandboxes · RL · GPU · exec API

Sandboxes / every shape

Every kind of sandbox.
One API.

Sandboxes for a single tool-call, workspaces that pause and resume across long agent sessions, environments built for RL rollouts, GPU sandboxes on the way, and an execution API with the Python data stack already installed. Need something we don't list — we'll build it with you.

create sandbox ↗ read the docs

no sandbox — create one

< 5 ms

step latency · custom RL

~1 s

to a running machine

isolated

from your infrastructure

per-second

active compute

Live instances

4 active
sb-a9f2 python running 38ms
sb-c4e1 node running 12ms
sb-7b3d python complete 91ms
sb-f21a bash spawning —

The catalogue

Six ways to get a machine.
Same API, same teardown.

01
on demand

Sandboxes

An isolated machine for agent tool-calls and one-off runs. Starts in about a second and bills only active seconds.

start
~1 s
isolation
per sandbox
billing
per second after 1 hr*
create one
02
pause · resume

Persistent workspaces

Pause a sandbox and its CPU and memory are released while the files stay. Resume later in the same workspace — for agent sessions that run across days, not minutes.

on pause
compute released
files
kept
on resume
same workspace
read the docs
03
agent training

RL environments

Run agent tasks in isolated sandboxes, check the outcome with your own verifier, and record the run.

task
your code and data
reward
your verifier
output
session and result
learn more
04
built with you

Custom RL sandboxes

Your simulator, your image, your hardware profile — co-designed with us and pinned to reserved capacity. Sub-5 ms steps, custom observation plumbing, whatever the training loop needs.

step latency
< 5 ms
image
yours
capacity
reserved
talk to us →
05
accelerated

GPU sandboxes

A sandbox with a GPU attached, for inference, fine-tuning runs and the CUDA code an agent writes. Created and torn down through the same API as every other sandbox.

hardware
GPU attached
workloads
inference · training
api
same as sandboxes
ask for early access
06
judge0-style

Execution API

Send code, get the result back. Judge0-style execution with NumPy, SciPy and pandas already resident, so nothing waits on an install.

numpy · pandas
preloaded
returns
output · time · memory
sandbox
managed for you
read the docs

Command line

One command,
then it's just your terminal.

The installer detects your OS and CPU, verifies a checksum and drops the boltz binary on your PATH. macOS and Linux, Intel and ARM; Windows works under WSL.

read the docs
shell 
curl -fsSL https://boltzlabs.cloud/cli/install.sh | sh
boltz auth paste an API key
boltz create pick a machine, interactively
boltz ls list your sandboxes
boltz connect <id> open an interactive shell
boltz exec <id> <cmd> run one command
boltz rm <id> destroy a sandbox

Tell us the shape

Want a sandbox built for
reinforcement learning?

talk to us →
Per-second active billingRoot accessEditor or terminalCustom RL on request