kvcache-ai/AgentENV

Features

• MicroVM Agent Sandboxes - Runs many isolated Firecracker microVM environments in parallel across machines so AI agents can execute untrusted or heavy workloads safely.
• Distributed Workload Orchestration - Spreads sandbox workloads over a cluster, with a gateway and scheduler that place, manage, and reach environments across many nodes.
• E2B-Compatible APIs - Serves an E2B-compatible HTTP API so existing E2B Python and TypeScript SDK code works unchanged against self-hosted infrastructure.
• Sandbox Cluster Schedulers - Provides a gateway and scheduler control plane that places and manages sandbox environments across a multi-node cluster of machines.
• Sandbox Forking - Captures incremental memory and filesystem snapshots, stores them in object storage or shared filesystem, and forks running environments into independent copies.
• Snapshot Fork Restores - Pauses, resumes, snapshots, and forks running environments so parallel experiments branch from a shared starting state and idle cost stays low.
• MicroVM Sandboxes - Runs each workload in its own Firecracker microVM with a dedicated Linux kernel so many environments run side by side without interfering.
• MicroVM Runtime Isolations - Runs each workload in its own Firecracker microVM with a dedicated Linux kernel so many environments share a host safely.
• Sandbox Templates - Builds reusable sandbox templates from OCI images and snapshots so teams launch new environments from a consistent baseline.
• Centralized Control Planes - Provides a central gateway and scheduler that place and manage sandbox environments across a multi-node cluster.
• On-Demand Layer Loading - Streams OCI image layers on demand through layered copy-on-write block devices, using local disk as a bounded cache of hot data.
• Sandbox Pause and Resume Controls - Saves memory and disk state of idle environments to free CPU and memory, then restores them within tens of milliseconds when new work arrives.
• Sandbox Cluster Scaling - Distributes large numbers of environments over many machines using a gateway and scheduler that place and manage sandboxes across the cluster.
• Warm Snapshot Restores - Restores environments from saved memory and disk state within tens of milliseconds, supporting pause, resume, snapshot and fork operations.
• Node Virtualization Modes - Runs each node in exactly one virtualization mode, KVM or PVM, provisioning only the matching Firecracker and guest kernel files.
• Copy-on-Write Root Layering - Streams OCI image layers into stacked copy-on-write block devices so many environments share one base image with private writes.
• Incremental Snapshotting - Captures only changed memory pages and disk blocks at each snapshot, enabling fast pause, resume, and forking of environments.

Star history