Back to products
Beta

Citadea SandBox

Ephemeral, isolated environments for AI agents

A safe place for agents to run code they wrote and data they fetched — destroyed the moment they are done.

Overview

Isolated compute pods designed for AI agents to execute generated code or parse external data with strict network isolation and auto-destruction after execution. Untrusted code runs safely, sandboxed web access included, every action audited.

Isolation you can trust with untrusted code

Each SandBox is a strictly network-isolated pod that auto-destructs after execution. Agents can run generated code or parse external data with sandboxed web access, and every action is recorded in a full audit trail.

Compared to Modal and E2B sandboxes.

What's included

  • Strict network isolation
  • Auto-destruction after execution
  • Untrusted code execution
  • Sandboxed web access
  • Native MCP integration
  • Full action audit trail

Quickstart

$ citadea sandbox create --agent <name>

Benchmarks

< 10 ms
Provision time
L3/L4
Isolation

Use cases

Agent-generated codeSandboxed web scrapingSecurity testingUntrusted execution

Integrations

MCPLangChainAutoGenCitadea CLI

Frequently asked