Open-source agent sandbox runtime

Cloud Sandboxes for AI Agents.

Give every agent a secure computer with lasting memory.

Designed for coding agents through MCP

CodexClaude CodeCursorMCP

Plan sandboxes for any agent workload.

From one research task to an enterprise agent fleet, each workload gets its own image, network access, resource limits, and requested artifacts.

semiconductor filings45 min
arxiv.org12 papers
sec.gov8 filings
sources.jsonrequested output

Deep research

Allow only the sources a research agent needs, cap its runtime, and request a structured report.

repo / checkout

$ cargo test

28 passed · 0 failed

+ allow api.github.com

- inherit host access

artifact → git.diff

Coding agents

Mount one repository, run its tests, and return the diff without exposing the rest of the host.

https://allowed.example

Browser

Files

Console

host allowed

Browser agents

Choose a browser-ready OCI image, specific websites, and a fixed memory ceiling.

Job queue4 CPU / 2 hr
index-repoplanned
run-testswaiting
write-reportwaiting
logs requested as run.log

Background jobs

Bound longer agent tasks with explicit compute, wall-clock time, logs, and cleanup expectations.

Infrastructure for enterprise fleets. Plan thousands of agent sandboxes.

Regional fleets

Plan capacity across regions without sharing files or credentials between sandboxes.

GPU-backed work

Describe GPU, memory, queue, and time requirements for training, evaluation, and inference.

Team controls

Give each team explicit compute limits, network access, logs, and requested outputs.

Set CPU, files, network, and time.

sandbox ready
commandcargo testfiles/workspaceoutputgit.diff

Files and commands

Choose the files and command each sandbox can use.

resourcesper sandbox
Cpu2 cores
Memory512 MB
Time20 min

CPU and memory

Set compute, memory, and a hard time limit for every run.

runtimeauto
macOSApple VZ
isolated from the host

Runtime

Match macOS, Linux, or Windows to an isolated sandbox runtime.

Preview a sandbox before it runs.

Doro checks the configuration and selects a runtime. It does not start the sandbox yet.

Available

Configuration checks and runtime selection.

Next

One isolated executor at a time.

Source available

Help build the safer default for agent work.