An open-source (MIT) sandbox platform released by KVCache.AI. It exists to hand AI agents a disposable Linux environment, spinning up kernel-level isolated environments in large numbers and at high speed using Firecracker microVMs. It was originally built to support agentic reinforcement learning (RL) training for large language models, and its defining characteristic is speed: booting or resuming from a snapshot takes under 50 milliseconds, and taking a snapshot takes under 100 milliseconds. Because it exposes an E2B-compatible HTTP API, existing E2B SDK code can be pointed at it almost as-is. The first version was published in July 2026, and the CLI runs on Mac and Linux.
Key Features
- Kernel-level isolation via Firecracker: Each sandbox runs as a microVM with its own Linux kernel. Isolation is stronger than shared-kernel container approaches, which suits workloads where an agent executes arbitrary code
- 50ms-class boot, resume, and snapshots: Memory and filesystem snapshots let an environment boot or resume in under 50 milliseconds, while capturing a snapshot takes under 100 milliseconds. Pausing environments and parking them becomes practical at a realistic cost
- Forking a running environment (up to 16 in parallel): A live environment can be cloned into as many as 16 child sandboxes. This pairs well with RL and tree-search style agent execution, where multiple actions are tried from the same state
- E2B-compatible API: SDKs and clients written against E2B work by changing only the endpoint. Moving from usage-based cloud pricing to your own infrastructure requires little rewriting
- Large-scale image handling with overlaybd and ublk: It combines overlaybd for on-demand OCI image loading with ublk for high-performance I/O, sharing the host page cache. The developers cite figures of 1.5 million images and a 9.6x memory overcommit ratio in production
- Snapshot persistence and distributed deployment: Snapshots can be stored in S3-compatible object storage or a shared distributed file system, and a gateway plus scheduler supports multi-node configurations
- Reverse proxy to services inside a sandbox: HTTP and WebSocket services started inside a sandbox are reachable from outside, which is useful for checking a development server an agent has launched
Pricing
| Plan | Price | Key Features |
|---|---|---|
| Open source (self-hosted) | Free (MIT license) | Full functionality on your own servers. The real cost is the machines, storage, and bandwidth you provide |
Pricing is current as of August 2026. AgentENV is self-hosted open-source software, and no official paid hosting plan has been announced. Please check the official repository for the latest information.
Pros & Cons
✅ Pros
- Boot and resume take under 50 milliseconds while retaining the strong isolation of microVMs
- Running environments can be forked, so branching execution from an identical state scales cheaply
- The E2B-compatible API means existing clients need almost no rewriting
- MIT-licensed and self-hostable, so you are not tied to usage-based sandbox billing
- Multiple deployment paths are provided: Docker Compose, Kubernetes, and manual setup
⚠️ Cons
- The execution host must be Linux (kernel 6.8 or later with access to
/dev/kvm). What runs on Mac and Linux is the CLI client - Environments without KVM require the PVM procedure, which raises the difficulty of setup
- Being self-hosted, the burden of procuring, monitoring, and updating servers falls on you
- The project is still on the 0.1 series, a stage where specification changes and bug fixes continue
- This is not a no-code managed service; infrastructure knowledge is assumed
Comparison with Similar Services
| Criteria | AgentENV | E2B | Daytona | Modal Sandboxes |
|---|---|---|---|---|
| Delivery model | Self-hosted (OSS) | Cloud (OSS edition available) | Cloud / self-hosted | Cloud |
| Isolation | Firecracker microVM | microVM | Container / VM | Container-based sandbox |
| Main use case | Agent training, large-scale parallel execution | Code interpreters, agent execution | Development environments, agent execution | General-purpose serverless execution |
| API compatibility | E2B-compatible | E2B (the original) | Proprietary SDK | Proprietary SDK |
| Pricing | Free (cost of your own infrastructure) | Usage-based | Usage-based | Usage-based |
Who Is It For
- Research and development teams that need to run large numbers of sandboxes in parallel for reinforcement learning or agent evaluation
- Developers implementing search-style agents that try multiple branches from the same state
- Product teams that have used E2B but want to move to their own infrastructure for cost or scale reasons
- Organizations with security requirements that demand stronger isolation than containers because agents execute arbitrary code
- People who can operate Linux servers themselves and are comfortable tuning infrastructure
Summary
AgentENV is a self-hosted runtime committed to one thing: provisioning the disposable environments handed to agents quickly, in bulk, and with strong isolation. The 50-millisecond-class boot and resume, together with forking of live environments, pay off in workloads that consume environments heavily, such as agentic RL and exploratory agent execution. On the other hand, the execution host assumes Linux plus KVM, and operations are on you. If you want the convenience of a managed offering, cloud services such as E2B fit better; if you work at scale and want to control cost and isolation yourself, this is a strong option.