Zero Trust AI for controlled sandbox automation
Brass tacks.
Global stacks.
Infrastructure automation for sandboxes on the systems you already run
Execution model
Use the capacity you own first. Add the cloud when it earns its keep.
GlobalStacks keeps orchestration centralized while sandboxes run near your code, services, data, and existing infrastructure. That gives agent workloads better runtime economics, lower latency, and stronger operational control.
Platform features
Operational control across your fleet
Cross-agent mesh
Agents form a private mesh for clustering, placement, health, and control-plane coordination across hosts.
Platform foundations
The same control plane, even when the work runs somewhere else.
Self-hosted clustering
Treat your own machines as a managed sandbox fleet.
Group hosts into clusters and place sandboxes by host, traits, tags, architecture, and runtime health without giving up control of the underlying capacity.
Shared sandbox storage
Durable files without hand-wiring host disks.
Attach shared sandbox volumes backed by JuiceFS so workloads can keep state, move between hosts, and use consistent paths across repeatable runtime sessions.
Windows operator path
Run the same operations from PowerShell.
Install gstacks, connect Windows hosts, and use current-user PATH and completion setup so Windows machines can participate in the same operational workflow.
ARM64 and Apple Silicon
Use modern local hardware where it fits.
Run gstacks and gstacks-agent on Linux ARM64 and macOS ARM64 hosts, including Apple Silicon machines used for development, testing, or local capacity.
Cost-aware placement
Make economics part of the runtime decision.
Place workloads where they make financial sense, keep owned capacity busy first, and reduce exposure to cloud bill shock when burst capacity is not required.
Cloud migration path
Add managed clusters later.
Keep sandbox execution consistent as infrastructure shifts across self-hosted machines and cloud regions, without rewriting the developer or agent workflow.
Local port forwarding
Bring one sandbox service back to your workstation.
Join a sandbox network from your workstation and forward a service alias to localhost for development, debugging, and integration checks.
Blueprint supply chain
Repeatable sandbox builds from trusted artifacts.
Import an OCI image, capture the process defaults, activate the blueprint, then launch sandboxes from the same named artifact every time. Blueprint records keep the source image and artifact reference visible so teams can reason about what is running.
OCI source
ghcr.io/acme/dev:sha256...
Blueprint
node-dev / active
Registry
built-in or connected private registry
Sandbox
repeatable rootfs + process defaults
Registry options
Use the built-in registry for GlobalStacks-managed blueprint artifacts, or connect Docker Hub, GHCR, ECR, self-hosted registries, and other private OCI registries with stored credentials.
Brokered authentication
Keep durable credentials out of sandbox workloads.
GlobalStacks keeps durable credentials in the trusted control plane and uses connected agents to broker the narrow operation a sandbox needs. The sandbox gets destination-scoped network reachability, injected headers only when HTTPS/TLS policy allows it, or a typed stream, not a pile of reusable cloud tokens.
Control-plane tokens stay outside
OAuth sessions and source-control credentials remain in the control plane or brokered secret store instead of being copied into sandbox files, logs, command arguments, or environment variables.
Access brokers outbound headers
Access policies can inject approved HTTP headers from write-only secrets through the agent-managed egress path, so application code does not need to hold the long-lived credential.
Secure network access is mediated
HTTPS and TLS egress can stay opaque for destination allow or deny rules. Header injection requires explicit TLS interception with a GlobalStacks-controlled CA, and ineligible or pinned clients fail closed.
Short-lived operational tokens
Sandbox ssh gateway, terminal, port-forward, registry, and agent stream paths use scoped short-lived credentials or hashed session tokens for the specific operation.
Zero Trust AI boundaries
Agentic workflows run in isolated sandboxes where access policies, brokered credentials, network reachability, storage mounts, and typed operations define exactly what AI can access or do.
Data plane
A data plane that works with you.
GlobalStacks separates where work runs from where durable state lives. Cluster storage, S3-compatible object stores, Cloudflare R2, and RustFS extensions give sandboxes portable files without turning provider credentials into sandbox secrets.
Cluster-aware storage
Attach durable storage to the cluster that owns the workload, then let agents mount the right volume backend when a sandbox starts.
S3 and R2 backed volumes
Cluster storage can use S3-compatible backends such as Cloudflare R2, with per-volume prefixes for chunks, metadata, and backups.
RustFS runtime extension
Run RustFS as a first-party data-plane object store on selected hosts through declared extension operations and typed broker calls.
Sync without credential sprawl
Agents receive scoped storage credentials for the operation they perform, while sandboxes consume mounted filesystems instead of provider keys.
Extensibility
Extend the console without trusting every extension.
GlobalStacks extensions are manifest-driven. The trusted console owns placement, rendering, consent, and operation authorization, while extension code runs in sandboxed UI and runtime boundaries.
Supported source control
Source control connections
Connect GitHub or Gitea repositories for Git-backed blueprint builds, extension runtime builds, and sandbox bootstrap flows.
Extension UI library
Author views with the GlobalStacks SDK and component contracts for forms, tables, data grids, status, progress, actions, and root views.
Sandboxed runtime protocol
Extensions declare a manifest, run in an isolated runtime, receive JSON operation input, and return JSON results through the broker.
Write your own extensions
Ship private or published extensions with declared permissions, egress, viewports, operations, commands, and runtime artifacts.
What can be extended
One control plane.
Any infrastructure.
GlobalStacks abstracts vendor-specific complexity. Whether work lands on an on-premises Proxmox node or cloud capacity in us-east-1, the sandbox execution model stays consistent.
- Unified sandbox management
- Distributed sandbox volumes
- Cross-agent Tailscale-compatible mesh
- Self-hosted clusters
- Linux ARM64 and Apple Silicon agents
- $1 per self-hosted cluster
- Centralized logging & metrics
- Infrastructure-agnostic APIs
target:
type: proxmox-lxc
node: 'pve-01'
sandbox:
name: 'agent-runner'
image: 'node:20-alpine'
resources:
cpu: 2
memory: 2048