[ Open-source · Built on Firecracker ]

Persistent & Secure
sandboxes for your agents.

Give your agents lightning-fast sandboxes with persistent state and versioned filesystems.

[01]FirecrackerHardware-Level Isolation
[02]<200msStartup
[03]VersionedFilesystem
[04]InfiniteSession Length
[01] Use cases4 patterns

Your agents need their own computers.

Fast, elastic, stateful infrastructure built for AI workloads.

[A] · Agent in a sandbox

Agent harnesses

  • Run the agent inside isolated, stateful sandboxes.
  • Give the agent full access to files, packages, bash, working state.
  • For research agents, personal agents, or long-running sessions.
Agent harnesses
// agent.ts — long-running harnessimport { Sandbox } from "@superserve/sdk"const sbx = await Sandbox.create({ name: "claude-sandbox" })await sbx.commands.run("npm i -g @anthropic-ai/claude-agent-sdk")await sbx.commands.run("claude -p 'refactor src/**/*.ts'")await sbx.pause() // resume later with the same state
[B] · Sandboxed tools

Isolated code execution

  • Execute AI-generated code, browsers, or system tasks in a sandbox.
  • Keep the harness outside, spin up sandboxes for tool calls.
  • For tool-calling agents that should not run untrusted code inside the harness itself.
Isolated code execution
// tool.ts — per-call sandbox for untrusted codeimport { Sandbox } from "@superserve/sdk"const sbx = await Sandbox.create({ name: "tool-call", timeoutSeconds: 300 })const out = await sbx.commands.run(  `python -c ${JSON.stringify(toolCall.input.code)}`)console.log(out.stdout)
[C] · Review and experimentation

Parallel agent swarms

  • Fork a workspace, run parallel agents in isolated sandboxes, pick the best result.
  • Snapshot the entire filesystem at any step, branch out to many environments, then resume from the exact state.
  • For agent swarms that modify the same files, or agents that need human reviews on multiple paths.
Parallel agent swarms
// swarm.ts — fan out 8 agents from a shared templateimport { Sandbox } from "@superserve/sdk"const results = await Promise.all(  Array.from({ length: 8 }, async (_, i) => {    const sbx = await Sandbox.create({      name: `agent-${i}`,      fromTemplate: "superserve/node-22",    })    return sbx.commands.run("agent run")  }))
[02] FilesystemVersioned · Fast

The fastest filesystem with built-in version control

Persistentauto sync to storage

vs. lost state at sandbox termination

<50msp95 random read on 10GB file

vs. multiple seconds with S3

<1stime to mount 10GB repo

vs. git clone taking minutes

Unlimitedconcurrent writers

vs. file locks on other durable filesystems

[03] SecurityBy default

Security built for AI workloads

Firecracker isolation

Every agent and tool call runs in its own microVM. No shared kernels or cross-tenant states. Stronger isolation than docker containers and local worktrees.

Credentials broker

Agents should not possess credentials. Our broker sits between your agent and the APIs it calls, eliminating credential exfiltration risk with brokered access.

Open source transparency

Verify every line of code. No black boxes. No hidden logic.

Network Controls

Control egress traffic to specified URLs and IPs.

superserve-ai/superserve
[04] Programmatic controlSDK · API

Programmatic control built for agents

Process execution

Execute code and commands in isolated environments with real-time output streaming.

File system operations

Manage file CRUD operations and checkpoint/rollback state programmatically.

Elastic resource allocation

Let your agents request CPU/memory resources on the fly from within the box.

[05] Pricing

Start for free, pay as you scale.

No credit card to start. Per-second billing.

Compute · vCPU$0.0504/ hour
Memory · GiB$0.0162/ hour
Storage · GiB$0.000108/ hour
Get startedStart freeNo card required

Your first sandbox
in seconds.

npm install @superserve/sdk
No credit card requiredOpen source131BYOC available