What is Superserve?
Superserve provides persistent, secure sandboxes for AI agents, built on Firecracker microVMs for hardware-level isolation. Agents get lightning-fast environments with versioned filesystems, sub-200ms startup, and infinite session length. The platform features a credentials broker to eliminate credential exfiltration risk, network controls, and open-source transparency.
Who it's for
- AI researchers who need isolated, stateful environments for long-running agent experiments
- Developer tool builders creating agent harnesses that require secure code execution
- Platform engineers managing parallel agent swarms that need forkable, versioned workspaces
Key features
Firecracker hardware-level isolation
Every agent and tool call runs in its own microVM with no shared kernels or cross-tenant states. This provides stronger isolation than Docker containers or local worktrees, making it suitable for running untrusted AI-generated code.
Versioned filesystem with built-in version control
The filesystem offers persistent auto-sync to storage with sub-50ms p95 random reads on 10GB files and under 1 second to mount a 10GB repo. Unlimited concurrent writers eliminate the file locks found on other durable filesystems.
Credentials broker for zero-trust security
The broker sits between your agent and the APIs it calls, eliminating credential exfiltration risk. Agents never possess credentials directly, only brokered access to the services they need.
Programmatic control via SDK and API
Execute code and commands in isolated environments with real-time output streaming. Manage file CRUD operations and checkpoint/rollback state programmatically, while agents can request CPU and memory resources on the fly.
What stands out
Give your agents lightning-fast sandboxes with persistent state and versioned filesystems.
This combination of persistence and speed is rare in the sandbox space. Most solutions either sacrifice startup time for statefulness or lose state when the session ends. Superserve delivers both sub-200ms startup and infinite session length with versioned filesystems, enabling agent workflows that were previously impractical.
Worth checking out if…
You're building AI agents that need secure, stateful environments with hardware-level isolation. The per-second billing (vCPU at $0.0504/hour, memory at $0.0162/GiB/hour, storage at $0.000108/GiB/hour) makes it cost-effective for both development and production use. The open-source transparency and Firecracker foundation provide confidence for security-conscious teams.
