# Sandbox Providers

> Alibaba Cloud Agent Sandbox lists the lowest managed rate: $0.0137 per vCPU-hour and $0.0069 per GiB-hour (Economy class), so a 2 vCPU / 4 GiB sandbox costs about $0.055 an hour. Northflank follows at $0.0167 and $0.0083, about $0.07 an hour. E2B, Daytona and Hopx all charge $0.0504 and $0.0162, about $0.17 an hour for the same box. Vercel, Cloudflare, Tensorlake, Upstash and AWS AgentCore bill only active CPU, which is cheaper for agents that sit idle between tool calls. For untrusted, model-written code, pick a microVM: it gives every sandbox its own kernel.

Every AI code sandbox, priced per vCPU-hour, with the source one click away. 31 providers, each field linked to its public source, last verified 2026-11-09.
When citing, link the canonical HTML page.

## Data
- [AI code sandboxes for agents, priced per vCPU-hour](https://sandboxproviders.com/): the full comparison table
- [Methodology](https://sandboxproviders.com/methodology/): what each field means and how it is verified

## Providers
- [Agent Sandbox (Kubernetes SIG)](https://sandboxproviders.com/kubernetes-agent-sandbox/): Kubernetes-native controller and CRDs for agent sandboxes, delegating isolation to gVisor or Kata Containers, with warm pools for fast allocation. For teams that run their own Kubernetes clusters.
- [Agent Sandbox on GKE (Google Cloud)](https://sandboxproviders.com/gke-agent-sandbox/): Google Kubernetes Engine's managed path for the open-source Agent Sandbox, running AI-generated code in gVisor-isolated pods with warm pools and a Python client. For teams on Google Cloud.
- [Alibaba Cloud Agent Sandbox](https://sandboxproviders.com/alibaba-agent-sandbox/): Alibaba Cloud runs each agent sandbox in its own VM, works with the E2B SDK, hibernates idle sandboxes without CPU charges and bills per second. For teams that want E2B-style sandboxes at the lowest list rate, or a Kubernetes API.
- [Amazon Bedrock AgentCore Code Interpreter](https://sandboxproviders.com/aws-agentcore-code-interpreter/): AWS's managed code-execution tool for agents, with pre-built Python, JavaScript and TypeScript runtimes, S3 access and CloudTrail logging. For teams building agents on AWS.
- [Azure Container Apps Dynamic Sessions](https://sandboxproviders.com/azure-dynamic-sessions/): Azure's Hyper-V-isolated sessions for running LLM-generated code, from prewarmed pools, as a built-in code interpreter or your own container. For teams building agents on Azure.
- [Beam](https://sandboxproviders.com/beam/): Beam is a serverless GPU and CPU cloud with a Sandboxes product for untrusted code, built on its open-source beta9 engine. For teams that want sandboxes and GPU inference on one bill.
- [Blaxel](https://sandboxproviders.com/blaxel/): Blaxel runs agent sandboxes as microVMs that drop to standby after a few seconds idle and resume in about 25 ms with memory intact. For agents with bursty, interactive workloads.
- [Claude Code Execution Tool (Anthropic)](https://sandboxproviders.com/anthropic-code-execution/): Anthropic's server-side sandbox lets Claude run Python and Bash inside the API call: a 1 CPU / 5 GiB container with no internet access that you can reuse for 30 days. For agents built on Claude that need data analysis and file generation without running their own sandbox.
- [Cloudflare Sandbox SDK / Containers](https://sandboxproviders.com/cloudflare-sandbox/): Cloudflare's Sandbox SDK starts container instances, each a microVM with its own kernel and network, controlled from a Worker. For teams already building on Cloudflare Workers.
- [CodeSandbox SDK](https://sandboxproviders.com/codesandbox-sdk/): The CodeSandbox SDK creates Firecracker microVM sandboxes that can be cloned from templates and resumed from snapshots. Together AI now documents it as its legacy sandbox SDK.
- [Daytona](https://sandboxproviders.com/daytona/): Daytona runs sandboxes for AI-generated code with a dedicated kernel, filesystem and network per sandbox, plus Linux VM, Windows and GPU options. Built for agent platforms that need fast creation and long-lived sandboxes.
- [Docker Sandboxes](https://sandboxproviders.com/docker-sandboxes/): Docker's sbx CLI runs AI coding agents in isolated microVM sandboxes on your own machine or on Docker-managed cloud compute. For developers who want to let coding agents run unattended without exposing the host.
- [E2B](https://sandboxproviders.com/e2b/): Cloud sandboxes for AI agents: each sandbox is a Firecracker microVM that agents drive through Python and JS/TS SDKs. Built for teams shipping coding agents, data-analysis agents and computer-use agents.
- [Firecracker](https://sandboxproviders.com/firecracker/): Firecracker is the open-source microVM monitor AWS built for Lambda and Fargate, and the isolation layer under several sandbox vendors. For platform teams building their own sandbox service.
- [Fly.io Sprites](https://sandboxproviders.com/fly-sprites/): Sprites are hardware-isolated Linux computers for AI agents on Fly.io, with a persistent 100 GB disk, fast checkpoints and no compute charge while asleep. For long-lived agent workspaces.
- [Freestyle](https://sandboxproviders.com/freestyle/): Freestyle provides full Linux VMs for AI agents with nested virtualization, hibernation with memory intact and forking. For app-builder and coding-agent products that need real VMs.
- [Google Agent Platform Code Execution](https://sandboxproviders.com/google-agent-platform-code-execution/): Google's managed code-execution sandbox for agents (formerly Vertex AI Agent Engine): sandboxes start and run code in under a second and keep their state for up to 14 days, with VPC Service Controls, CMEK and HIPAA support. For teams building agents on Google Cloud.
- [Hopx](https://sandboxproviders.com/hopx/): Hopx runs untrusted and AI-generated code in microVM sandboxes that start in about 100 ms, billed per second. For agent builders who want E2B-style sandboxes with a large starting credit.
- [Koyeb Sandboxes](https://sandboxproviders.com/koyeb/): Koyeb Sandboxes are isolated environments for AI-generated and untrusted code on Koyeb's serverless platform, with scale-to-zero. For teams that want sandboxes next to their Koyeb apps and GPUs.
- [microsandbox](https://sandboxproviders.com/microsandbox/): Open-source, local-first microVM runtime and library for running untrusted workloads on your own machine or servers. For teams that want self-hosted hardware isolation with no vendor.
- [Modal Sandboxes](https://sandboxproviders.com/modal/): Modal's Sandboxes API runs untrusted code in gVisor-isolated containers on Modal's serverless cloud, with GPU access. For Python-first teams that already use Modal or need GPUs next to agent code.
- [Morph Cloud](https://sandboxproviders.com/morph-cloud/): Morph Cloud runs MorphVM instances and devboxes that can be paused, resumed, snapshotted and branched. For agent and RL workloads that explore many paths from one state.
- [Northflank Sandboxes](https://sandboxproviders.com/northflank/): Northflank runs each sandbox in its own Cloud Hypervisor microVM (gVisor for GPU), on Northflank's cloud or inside your own AWS, GCP, Azure or Oracle account. For teams that want no runtime limit and BYOC.
- [OpenAI Containers (Code Interpreter / Hosted Shell)](https://sandboxproviders.com/openai-containers/): OpenAI's hosted containers give models in the Responses API a sandboxed VM for Python and shell, in 1, 4, 16 or 64 GB sizes, created automatically or through the Containers API. For agents built on OpenAI models.
- [Prime Intellect Sandboxes](https://sandboxproviders.com/prime-intellect-sandboxes/): Prime Intellect runs each sandbox in a hardware-isolated microVM, with pause and resume that keeps memory and processes, checkpoints you can fork, and thousands of sandboxes starting at once. Built for RL rollouts, evals and agent fleets.
- [Riza](https://sandboxproviders.com/riza/): Riza is a code-execution API that runs LLM-generated Python, JavaScript, TypeScript, Ruby and PHP in WebAssembly isolation. For short tool calls rather than long-running environments.
- [Runloop](https://sandboxproviders.com/runloop/): Runloop Devboxes are VM-isolated sandboxes for AI coding agents, with blueprints, snapshots and suspend/resume. Aimed at teams building and evaluating coding agents.
- [Tensorlake Sandboxes](https://sandboxproviders.com/tensorlake/): Tensorlake runs sandboxes on Firecracker microVMs, billed on active CPU with a 20% floor, with snapshot and resume on Pro. For agent and document-processing workloads.
- [Together Code Sandbox](https://sandboxproviders.com/together-code-sandbox/): Together AI's code sandboxes for building AI development environments, with hibernation, VM cloning and git-versioned storage. For teams already using Together's inference platform.
- [Upstash Box](https://sandboxproviders.com/upstash-box/): Upstash Box gives agents persistent sandboxes billed on active CPU only, with memory free and boxes that freeze when idle. For agents that idle most of the time.
- [Vercel Sandbox](https://sandboxproviders.com/vercel-sandbox/): Vercel Sandbox runs agent-generated code in Firecracker microVMs and bills CPU only while code is actively running. Suited to teams on Vercel and agents that spend most of their time waiting on LLM calls.

## Optional
- [All records in one file](https://sandboxproviders.com/llms-full.txt)
- [Sitemap](https://sandboxproviders.com/sitemap.xml)