Concept

Don’t limit the capability. Give it a safe boundary.

AI agents can combine code execution, file operations, SaaS integrations, and data processing to get real work done from a natural-language prompt. But in security- and compliance-sensitive environments, using powerful agents at all often gets restricted. GX Sandbox doesn’t clip that capability — it gives it a boundary where it can run safely.

The constraints of enterprise use

Agents running freely on a laptop are productive, but in an enterprise they raise real concerns:

Excessive access to local files and credentials.
Outbound network requests that are hard to control and audit.
Little organizational visibility into what runs — and what it produces — on personal machines.
User permissions and agent execution permissions aren’t separated, blurring accountability.
Hard to meet the evidence, permission-separation, and audit requirements of regulated industries.

An execution & control layer between your app and your data

At the core of GX Sandbox is microVM isolation — Firecracker-class. Orchestrated as a cluster and paired with middleware that controls each sandbox, it runs many agents safely, scoped per user and per task.

Layer	How GX Sandbox designs it
UI / application	Verify the user’s identity and declare the verified principal to GX Sandbox.
Execution layer (GX Sandbox)	Run bash, Python, Node.js and more inside a Firecracker-based microVM; inject permissions; govern and persist outbound connections, traffic, and execution logs within the boundary.
Data sources	Access SaaS such as Google Drive, Teams, and Slack — and internal data — only within the permissions bound to the principal.

Authorization centered on the principal

A “principal” is the subject an agent runs as — the identity of a user (or application) that your app has verified. Your app authenticates and authorizes, declares the verified principal to GX Sandbox, and GX Sandbox injects it into the boundary. The agent runs with only that user’s permissions — not a generic high-privilege identity.

Principal verification

Your app verifies the user’s identity and declares it to GX Sandbox as a verified subject.

Permission injection

The user’s permissions are injected into the sandbox as environment and execution context.

External connections

SaaS accounts and MCP/connectors (Google Drive, Teams, Slack, …) are loaded bound to the user’s permissions.

Boundary management

The sandbox boundary and the permission boundary are the same, making the agent’s execution accountable.

In one line: GX Sandbox is the platform to not just run AI agents, but to run them with permissions, safely, and auditably.

See how it works Start free