Superserve is an open-source sandbox platform built on Firecracker microVMs that provides persistent, stateful, and secure computing environments for long-running AI agents — capable of indefinite pauses, instant resumption, and parallel forking.
What is Superserve?
Superserve gives AI agents their own dedicated computer: a durable Firecracker microVM that pauses indefinitely, resumes instantly, and runs as long as the agent needs. It takes agents (or tool execution calls) as input and produces isolated, stateful sandboxes with full filesystem access, bash, packages, and network controls. The platform is open source (available on GitHub), built by the Superserve team, and integrates with harnesses like Claude Code, Codex, OpenCode, Kilo Code, Hermes, and OpenClaw.
Key Features
- Indefinite sessions — No 24-hour limit; sandboxes can run for days or weeks, pausing automatically to save idle costs and resuming exactly where they left off.
- Durable state — Snapshot, fork, and resume the entire VM at any point. Enables checkpointing, rollback, and parallel exploration across branches.
- Firecracker isolation — Each sandbox runs in its own microVM with no shared kernels or cross-tenant state, providing stronger isolation than containers.
- Credentials broker — A broker sits between the agent and external APIs, preventing direct credential exposure and eliminating exfiltration risk.
- Network controls — Restrict egress traffic to specified URLs and IPs; all outbound connections are logged and configurable by the user.
- Preview URLs — Expose ports from inside a sandbox via a public URL for instant previews of web apps or services.
- Docker support — Run Docker containers natively inside the sandbox with full tooling compatibility.
- Streaming & durable filesystem — Real-time output streaming from commands, and a filesystem that persists across sessions exactly as the agent left it.
Who is it for?
- AI agent developers — Run long-lived agents (e.g., coding assistants or research bots) inside stateful sandboxes that can pause and resume without losing context.
- Platform teams building agent harnesses — Integrate with MCP, TypeScript, or Python SDKs to provide sandboxed tool execution for AI-generated code, browsers, or system tasks.
- Security-conscious organizations — Use SOC 2 Type 2, HIPAA, and GDPR-compliant sandboxes with credential brokering and egress controls to safely run untrusted AI workloads.
- Engineering teams doing parallel experimentation — Fork a workspace into multiple isolated sandboxes, run agents in parallel, then pick the best result and resume from a previous snapshot.
What can you do with Superserve?
- Agent runtime: Run the agent harness (e.g., Claude Code) inside an isolated sandbox with full filesystem and network access, then pause the VM when idle and resume later without rebuilding state.
- Isolated code execution: Execute AI-generated code, browser automation, or system tasks in a sandbox while keeping the harness outside, spinning up and tearing down sandboxes per tool call.
- Multi-agent systems: Fork a workspace into many sandboxes, let parallel agents explore different approaches, then compare results and roll back to a clean snapshot.
How does Superserve work?
Sandboxes start instantly from customizable templates. The user can take a snapshot at any point to checkpoint, fork, or roll back. Sandboxes can be paused indefinitely (no idle cost) and resume automatically from the exact state the agent left. The SDK provides programmatic control via streaming, durable filesystem, elastic resource allocation, and Preview URLs.
Pricing
Superserve is free to start (no credit card required) with per-second billing. Compute is $0.0504/hour per vCPU, memory is $0.0162/hour per GiB, and storage is $0.000108/hour per GiB. BYOC (bring your own cloud) is also available.
FAQ
