PLANE_01 · SANDBOX / AGENT RUNTIME

A secure runtime where agents can actually do things.

Isolated microVMs, controlled tools and replayable traces — give agents code, browser and shell access without giving up control.

Runtime
agent sandbox
armed
TaskAnalyze & edit report
IsolationDedicated microVM
ToolsAllowlisted only
CheckpointBefore write-back
AuditReplayable trace
bash — sandbox provisioning

01$ sandbox.spawn(image="agent-base:2.4", cpu=2, mem=4G)

02$ tools.allow = [python, browser, fs:/workspace]

03$ network.egress = allowlist(zirax.ai, pypi.org)

04→ checkpoint.before("write_back") = armed

05# trace sbx-88214 · every call recorded · 90d

Runtime boundary

Agents act inside a signed boundary, not a free-running script.

Every ZIRAX sandbox is a disposable microVM: isolated from your network, limited to allowlisted tools and domains, and destroyed the moment the run ends.

microVM isolationtool allowlistegress controlimmutable trace
// workflow

From task to auditable outcome.

01 · DEFINE

Define environment

Pick an image, resources and lifetime for the run.

02 · GRANT

Grant tools

Allow only the APIs, files, packages and domains the task needs.

03 · EXECUTE

Execute isolated

Run with ephemeral storage and snapshot checkpoints.

04 · REVIEW

Review & replay

Approve sensitive actions; replay the full trace any time.

// capabilities

Built for agents that take action.

CAP_01

MicroVM isolation

Per-agent, per-run virtual machines — not shared containers. Spawns in ~120ms, destroyed on exit.

CAP_02

Code & browser tools

Python, shell, headless browser and file tools pre-installed, versioned and patched by ZIRAX.

CAP_03

Tool & network allowlists

Grant exactly the tools, paths, packages and egress domains a task requires — nothing else.

CAP_04

Checkpoints & approvals

Pause before irreversible actions. Human or rule-based sign-off resumes the run.

CAP_05

Replayable audit traces

Every command, tool call and state change recorded and replayable for review or compliance.

CAP_06

Elastic scale

Thousands of concurrent sandboxes with per-team quotas, budgets and automatic cleanup.

// runtime, by the numbers

Specs

~120ms

Cold sandbox spawn, ready for tool calls.

10k+

Concurrent sandboxes per region, elastic.

90 days

Default trace retention; configurable for audit.

Let your agents act — safely.

Talk to us about sandbox pilots, on-prem images and approval workflows for your team.

Request access →