Code & AI agent sandboxes: read the guide

Cube Sandbox (CubeSandbox)

Open source toolSandbox

Instant, Concurrent, Secure & Lightweight Sandbox Service for AI Agents

Homepage Docs Repository

Scanned 25 Sep 2026 · Schema v5 · features v19

Licence: Apache-2.0
Self-hosted; you pay for the infrastructure.
Free tier: Free to run
Isolation: Kata / gVisor
hardened or user-space kernel
Regions: —
Not disclosed on the public surface.
Capabilities: 17 / 57
10 partial · 23 fail · 4 not assessed
Agent readiness: 7.5/10 · 6.5/9
read · drive, pre-login surface

Advertised for

LLM generated code execution · Agentic RL rollouts (SWE bench) · Long running agent processes (OpenClaw, Hermes) · Browser automation · Ubuntu desktop / GUI automation · Service stacks (Redis, MySQL) with fast clone

Built for

Developers, then technical

· hero is benchmark numbers and architecture: RustVMM, KVM, eBPF, CoW memory forks
· onboarding is provisioning a KVM-capable Linux server and running an installer
· vocabulary is MicroVMs, templates, CubeVS, PVM kernels, Helm

01 Runtime Each fact carries its source; dots show confidence. Open a row for the evidence.

Isolation & lifecycle

Isolation primitive

Kata / gVisorhardened or user-space kernel

Persistence

Sandbox writes are ephemeral by default. Persistence comes from pause (memory + filesystem snapshot to disk; cross-node via S3, preview), S3-backed or JuiceFS Volumes with their own lifecycle, and host mounts from the compute node.

Snapshots / forks

Snapshot (memory + filesystem), clone into N independent copies, in-place rollback, create-from-snapshot; the CubeCoW engine does this at hundred-millisecond granularity.

Resources & limits

Storage

Template-derived rootfs with a writable layer (set per template); S3 Volumes (bundled MinIO, or external S3/COS/R2), a JuiceFS volume plugin (v0.7.2), host mounts, and a pluggable Volume framework.

Limits & quotas

Set by the operator's hardware. Idle timeout per sandbox (seconds; -1 = never; on_timeout kill or pause). Nodes stop accepting new sandboxes above 80% disk (error 130597). Claims: <60 ms cold start, <5 MB memory overhead per sandbox, thousands per node; a user-reported RL density of 80–100 1-vCPU/4 GiB sandboxes on a 128-core host.

GPU

none (roadmap: GPU Sandboxes)

Access

Deploy interface

apisdkclidashboardcontainer-imagetemplate

SDK languages

PythonGoJavaScript

Network & secrets

Egress controls

Per-sandbox allow_out/deny_out (IPv4/CIDR/DNS names with A-record learning) enforced in eBPF by CubeVS. Rules marked L7 route HTTP(S) to CubeEgress for domain/path/method filtering, credential injection and JSONL audit logs. Inbound public access can require a per-sandbox traffic token.

Secrets

Env vars at sandbox creation, or CubeEgress-held credentials injected as headers on matching outbound requests so the sandbox never sees them (used for AgentHub LLM keys).

Bring your own cloud

Placement & use

Jurisdiction

Advertised workloads

LLM generated code executionAgentic RL rollouts (SWE bench)Long running agent processes (OpenClaw, Hermes)Browser automationUbuntu desktop / GUI automationService stacks (Redis, MySQL) with fast clone

02 Capabilities 17 pass · 10 partial · 23 fail · 57 of 61 assessed

Supported — Documented on the public surface and verified.
Partially supported — Exists with caveats: gated, thin, or enterprise-only.
Not supported — Checked and not offered.
Not assessed — Scanned before this capability was tracked. Not a lack.

Compute & storage 5/10

Persistent diskStorage that survives restart/stop (not ephemeral-only).

Supported

Object storageManaged S3-compatible / blob object store the platform PROVIDES to a user's workloads (buckets, presigned URLs). DISTINCT from persistent-disk (block/volume attached to an instance) and from a self-host dependency (source.self_host.deps[]=["object-storage"] is an infra dep to RUN an OSS tool — the opposite direction). SCOPE varies and matters: native first-class (Railway Buckets, Cloudflare R2) vs a sibling SKU / extension (Fly Tigris, Bunny Storage, DigitalOcean Spaces) — record which in the evidence quote / note. Admitted v15.

Not supported

Snapshots / forkingSnapshot or copy-on-write fork of a running instance (filesystem-only counts). Whether the snapshot also restores memory + running processes is recorded separately by memory-snapshot (v19).

Supported

Live-state (memory) snapshotsA snapshot / checkpoint / pause captures a running instance's MEMORY and PROCESS state and restores it, so running processes continue where they left off instead of a freshly booted instance with the same files. Counts whether user-triggered (snapshot, checkpoint, pause-to-storage) or used to restore pre-initialised state (memory snapshots for fast cold starts). Distinct from snapshots-forks (any snapshot or fork, including filesystem-only; a product passing that key may still fail this one) and from scale-to-zero (an idle-cost property; suspension may or may not keep memory). Actor multiplexing (see scale-to-zero) records its live-state half here — only when the suspended actor's MEMORY/process state is what comes back, not a freshly booted worker with the actor's files.

Supported

Scale to zeroIdle instances suspend at no / negligible cost. ACTOR MULTIPLEXING (many suspended stateful actors sharing a small pool of pre-warmed workers, woken on request — Agent Substrate, AX, kagent; Rivet's hibernating actors) has NO key of its own (proposal REJECTED v20) — score the idle-cost half HERE and the live-state half under memory-snapshot; the oversubscription ratio / density is a performance characteristic the registry does not key.

Supported

GPU computeGPU-backed instances available.

Not supported

Multi-regionCompute placeable in more than one geographic region.

Not supported

Sandbox templates / custom imagesBuild reusable sandbox/VM images from an OCI image or Dockerfile (a base you fork instances from).

Supported

Cloud-init / ignition configCustomise a VM / microVM / instance at boot via cloud-init (or ignition) user-data — the same boot-metadata interface as AWS/Azure/GCP. Inject packages, users, SSH keys, files, and run-commands declaratively at launch. Distinct from sandbox-templates (a PREBUILT image you fork instances from) — this is PER-LAUNCH declarative customisation, not a baked image.

Not supported

Desktop / computer-use sandboxA virtual desktop (GUI) for LLM computer-use — browser + desktop automation, often via noVNC/VNC.

Partially supported

Selectable isolation runtimeThe user chooses the isolation runtime per workload / environment / sandbox from more than one option (e.g. microVM vs gVisor vs plain container, native namespaces vs gVisor, gVisor vs VM). Records the CHOICE; the isolation_primitive facet still records the best-evidenced primary. A platform that switches runtime automatically without a user choice (e.g. gVisor forced for GPU) does not pass on that alone.

Not assessed

Networking & delivery 3/8

Automatic HTTPSTLS endpoints provisioned automatically per workload.

Partially supported

Built-in reverse proxyPlatform routes external traffic to workload services.

Supported

Custom domainsBring-your-own domain via CNAME/DNS.

Partially supported

Public preview URLsShareable per-deploy / per-branch URLs.

Supported

Direct public IPv6Per-instance stable public IPv6 for raw TCP/UDP (beyond a managed HTTPS proxy) — e.g. mosh / raw sockets.

Not supported

Egress controlsConfigurable OUTBOUND network policy — egress firewall / allowlist / egress proxy on traffic LEAVING a workload. Scoped to OUTBOUND only (v15): private-network REACHABILITY between a user's own services is private-networking, not this. Static-egress-IP belongs here (an outbound concept).

Supported

Private networkingA platform-internal private network between a user's OWN services — service-to-service connectivity, VPC, service mesh, and/or service discovery, not exposed to the public internet. DISTINCT from egress-controls (OUTBOUND firewall/allowlist), reverse-proxy (PUBLIC inbound ingress), and direct-ipv6 (per-instance PUBLIC addressing). Admitted v15.

Not supported

Email send / receiveSend and/or receive email from the workload.

Not supported

Edge KV storeA globally-replicated key-value store read from WITHIN an edge runtime at the PoP (low-latency, eventually-consistent) — Cloudflare Workers KV, Deno KV, Fastly KV Store, Akamai EdgeKV, Vercel Edge Config. Distinct from first-party-db (a managed DATABASE service you provision and connect to) and from http-query-api (HTTP access to an external store) — edge-kv is colocated with the edge compute and read at the edge.

Not applicable

Developer access 6/9

Public APIDocumented programmatic API.

Supported

OpenAPI specMachine-readable OpenAPI / Swagger spec published.

Supported

CLIOfficial command-line tool.

Supported

SDKOfficial client library/libraries.

Supported

SSH accessDirect SSH into workloads.

Partially supported

Filesystem APIRemote file operations over the API (list/read/write/watch/copy/upload) without SSH.

Supported

Git deployDeploy from a Git repo / push-to-deploy.

Not supported

Private container registryHosted private image registry.

Partially supported

E2B-compatible APIDrop-in compatibility with the E2B sandbox API/SDKs — existing E2B client code runs against this product by changing the endpoint/domain + key, with no business-code changes. E2B itself passes as the reference implementation. Distinct from public-api / sdk (the product has its own API — this key says it also speaks another vendor's) and from self-host-oss (a self-hosted E2B-compatible server is scored on both).

Supported

HTTP data APIQuery / read / write the stored DATA over plain HTTP/REST (or GraphQL) WITHOUT the native protocol driver (Postgres wire / Redis RESP) — agent- and edge-friendly data access via fetch(). The DATA plane; distinct from public-api (a control-plane / management API). Auto-generated-from-schema REST/GraphQL (PostgREST, pg_graphql) is a flavour of this.

Not applicable

AI & agents 1/11

Built-in coding agentShips a first-party AI coding agent in the product.

Not supported

Agent skillsInstallable / declarable agent skills or tools.

Partially supported

AI app builderPrompt-to-app / vibe-coding generation surface.

Not supported

MCP serverExposes a Model Context Protocol server for agents.

Not supported

MCP clientActs as an MCP client / consumes MCP tools.

Not supported

MCP-server hostingRun/host MCP servers as a workload on the platform (distinct from exposing a first-party MCP control server — that is mcp-server).

Not supported

LLM gatewayRoutes LLM API calls through the platform (multi-provider).

Not supported

BYOK for LLMsBring-your-own LLM provider keys.

Partially supported

Model inference API (first-party)A FIRST-PARTY hosted LLM/model inference API — the provider serves its OWN catalogue of (usually open-weight) models behind a (typically OpenAI-compatible) endpoint, billed per token. Distinct from llm-gateway (a multi-PROVIDER proxy/router for YOUR calls to external providers like OpenAI/Anthropic) and from byok-llm (bring your own provider keys) — here the platform IS the inference provider, serving its own models.

Not supported

Autonomous QA agentA first-party AUTONOMOUS AI agent that QA-TESTS a running app / implementation — exercising it like a real user (clicking through flows, checking behaviour) and reporting bugs/defects (often back to a coding agent, or before a human review). Distinct from built-in-agent (a code-AUTHORING agent — this one VERIFIES the result, it does not write the feature; a product can ship both), from ai-guided-code-review (organises a diff for a HUMAN to review — not an autonomous tester), and from observability (runtime metrics/logs, not active testing).

Not supported

Credential-injection proxyThe platform holds real secrets OUTSIDE the workload and injects them on the wire (an egress / broker proxy swaps a placeholder or phantom token for the real credential on matching outbound requests), so code running in the sandbox — e.g. an untrusted agent — never sees or can exfiltrate the real key. Distinct from secrets-management (storage + injection INTO the workload as env vars/files — the workload holds the secret) and from egress-controls (outbound allow/deny POLICY; a product can have an allowlist without injecting credentials, and vice versa).

Supported

Preinstalled coding agentThe platform installs and launches a THIRD-PARTY coding agent (a harnesses.yml entry — Claude Code, Codex, OpenCode, Pi ...) inside a workspace/sandbox it provisions, with the agent's auth wired in (your own account or API key), so you start an agent session rather than an empty box. Distinct from built-in-agent (a FIRST-PARTY agent the vendor ships) and from harnesses[] (which records that a harness CAN target the product, not that the product runs it for you).

Not assessed

Data 0/1

First-party databasePlatform-provided managed database.

Not supported

Database branchingCopy-on-write branch of a DATABASE (schema + data) as a dev-lifecycle primitive — branch your data like code, usually instant via a metadata operation. Distinct from snapshots-forks (forks a running compute INSTANCE / VM state) and from preview-environments (a full per-PR environment that may USE a DB branch). For a data product a CoW branch maps HERE, not to snapshots-forks.

Not applicable

Vector searchFirst-class storage, indexing, and similarity search of vector embeddings — a native vector type/index or a managed vector DB. The data-layer primitive for RAG / semantic retrieval.

Not applicable

Ephemeral instant DB (no-signup)Provision a working, usable instance via a single command / HTTP call with NO account, signup, or card — typically time-boxed (expires unless claimed). A zero-onboarding path an agent can hit directly.

Not applicable

Platform & operations 0/5

Scheduled tasksCron / scheduled job primitive.

Not supported

Durable workflowsDurable execution / long-running workflow primitive.

Not supported

Preview environmentsA full EPHEMERAL environment spun up per PR / branch — compute + addons + (often) a DB branch — torn down on merge. More than preview-urls (which is just a shareable URL for one deploy).

Not supported

Built-in observabilityFirst-party metrics / logs / tracing / APM / profiling included in the platform (no third-party stack required to see what a workload is doing).

Partially supported

Managed CI/CD runnersPlatform-hosted, drop-in CI/CD runner VMs (GitHub Actions / GitLab), typically faster and cheaper than the CI vendor's own hosted runners, billed per runner-minute; isolated + ephemeral per job. Distinct from git-deploy (push-to-deploy of an APP) and scheduled-tasks (cron) — this is managed CI COMPUTE you point your existing pipelines at.

Not applicable

Project template marketplaceA marketplace / gallery of prebuilt PROJECT STARTER templates (apps or websites) a user clones or launches a new project from and then customises — the "start from a template" surface of a builder / PaaS. Distinct from sandbox-templates (a COMPUTE base IMAGE you fork runtime INSTANCES from — Dockerfile / OCI), from app-builder (the prompt-to-app GENERATION surface — a product can have both), and from one-click OSS-app catalogues / helm-addon-marketplace (deploy a FINISHED app as-is, not a starter you build on).

Not supported

Organisation & security 2/6

Built-in authPlatform-provided authentication / IAM for workloads.

Partially supported

SSO / SAMLEnterprise single sign-on.

Not supported

Secrets managementFirst-class secrets storage / injection.

Partially supported

TeamsMulti-user team accounts.

Not supported

SharingShare a running workload/resource via link (optionally gated).

Supported

BYOCRuns inside the customer's own cloud account.

Not applicable

Deploy into customers' VPCsA SaaS vendor uses the platform to deploy and operate ITS software inside ITS CUSTOMERS' cloud accounts / VPCs (nested or "customer BYOC"), managed from the vendor's control plane. Distinct from byoc (your own workloads in your own cloud account).

Not assessed

Workload identityThe platform gives each workload its own verifiable identity (OIDC token, SPIFFE SVID, cloud IAM federation) that it exchanges for short-lived credentials, so no long-lived static secret is stored or injected. Distinct from secrets-management (stores / injects static secrets) and credential-injection-proxy (swaps real secrets in on the wire).

Not assessed

Self-host / OSS coreOpen-source core that can be self-hosted.

Supported

Proposed, not verified

Capabilities the scan saw but the registry does not track yet. A key is admitted only once two distinct products exhibit it. These are not scored.

Cross-node pause/resume: suspend a sandbox on one node and resume it on another via an S3 backendcross-node-migration
“with an S3 backend, suspend a sandbox on one node and resume it on another, or create sandboxes from snapshots (preview)” cubesandbox.com
Transparent agent-tool redirection: host coding agent's Bash tool calls are rewritten by a hook to run inside a sandboxagent-tool-redirect-hook
“A PreToolUse hook intercepts the tool call itself, so Bash isolation is transparent and complete — there is no path for a command to skip it.” cubesandbox.com

03 Agent readiness Pre-login surface · counts, not a grade

Can an agent read it?

parse-ability · docs, metadata, machine-readable surfaces

7.5/10

6 full · 3 partial · 75% of applicable checks

llms.txt
Supported
“https://cubesandbox.com/llms.txt: 46 KB, links every guide, architecture page, blog post and changelog in English and Chinese.”
llms-full.txt
Supported
“https://cubesandbox.com/llms-full.txt: 2.6 MB full-docs dump (saved as a POINTER stub).”
robots.txt AI policy
Partially supported
“robots.txt: 'User-agent: * / Allow: / / Disallow: /404.html'. Generic allow-all; no AI UA named.”
JSON-LD
Not supported
“No application/ld+json in the homepage HTML.”
Sitemap
Supported
“sitemap.xml lists 256 absolute URLs and is referenced from robots.txt.”
Meta tags
Partially supported
“<title>CubeSandbox</title> and a descriptive meta description, but no og:/twitter: tags.”
Markdown negotiation
Partially supported
“Every docs page has a .md twin served as text/markdown (e.g. /guide/quickstart.md), but Accept: text/markdown on /guide/quickstart returns text/html.”
Content efficiency
Supported
“~4.8 KB readable text / 75.6 KB HTML ≈ 6.3%, server-rendered VitePress.”
Docs discoverable
Supported
“Header nav: Home · Guide · Architecture · Developer · Roadmap … reachable without JS.”
Machine-readable pricing
Not applicable
“No pricing: free Apache-2.0 self-host tool with no commercial tier on this domain.”
No anti-bot wall
Supported
“Static GitHub Pages behind Fastly; no challenge on any path fetched.”

Can an agent drive it?

act-ability · APIs, SDKs, MCP, credentials, webhooks

6.5/9

6 full · 1 partial · 72% of applicable checks

Public API
Supported
“E2B-compatible REST API (CubeAPI) documented by openapi.yml and the SDK guides.”
OpenAPI spec
Supported
“https://github.com/TencentCloud/CubeSandbox/blob/master/openapi.yml: valid OpenAPI 3.1.0, 26 paths.”
MCP server
Not supported
“No Model Context Protocol server in docs, llms.txt or the repo tree; /.well-known/mcp → 404.”
SDK packages
Supported
“Python (PyPI cubesandbox 0.7.0), Go and Node SDKs + stock E2B SDKs.”
Auth documented
Supported
“https://cubesandbox.com/guide/authentication: Bearer / X-API-Key credentials, callback request format, method+path validation warning, error codes.”
Self-serve credentials
Not applicable
“Self-hosted: API keys are whatever the operator's auth callback accepts; there is no issuing service.”
Webhooks
Not supported
“No webhook/event-callback feature documented for sandbox lifecycle events.”
CLI tool
Supported
“https://cubesandbox.com/guide/cli-tools: cubemastercli, cubeopscli, cubecli, cube-runtime.”
Error model
Supported
“openapi.yml responses reference components/schemas/ApiError; the v0.6.0 release post notes 'Network and security proxy aligned with E2B error code semantics'.”
Rate limits documented
Not applicable
“Self-hosted API with no documented rate limiting; capacity is the operator's hardware.”
Agent config files
Partially supported
“An OpenClaw SKILL.md (examples/openclaw-integration/skills/cube-sandbox) for agents that use Cube. The repo's AGENTS.md and .claude/ directory are contributor tooling, not product agent config; no /.well-known/agent.”

04 Coding harnesses Which agents and IDEs are documented to run here

Advertised integrations 5

Claude Code Anthropic
Supported
“This guide shows how to keep Claude Code running on your host while transparently redirecting every Bash command it runs into an isolated CubeSandbox MicroVM, using a PreToolUse hook.”
Source cubesandbox.com
OpenCode opencode
Supported
“Run OpenCode on your host, but execute every bash command it issues inside an isolated CubeSandbox MicroVM.”
Source cubesandbox.com
Pi earendil-works
Supported
“Run the Pi coding agent — a terminal-native AI coding agent — inside CubeSandbox MicroVMs.”
Source cubesandbox.com
OpenClaw Peter Steinberger
Supported
“The Digital Assistant (AgentHub) uses Cube Sandbox to create and manage OpenClaw assistants.”
Source cubesandbox.com
Hermes Agent Nous Research
Supported
“Hermes Agent: Running a Resident Agent Platform in CubeSandbox”
Source cubesandbox.com

Plausible compatibility 0

Derived, not stored: the product passes SSH access, so any CLI coding agent is installable inside it. This is plausible compatibility, not a supported integration.

No general shell access documented, so nothing is derived. Absence is not unsupported.

05 Compliance & regions Standards held, with ordinal rigor; never a score

Certifications held 0

No attestations found on the public surface. Empty means nothing is advertised, after checking, not that nothing was checked.

Regions —

No region list captured.

06 Open source Apache-2.0 · fetched by machine, not quoted from marketing

Licence & repository

Licence: Apache-2.0OSI
Primary language: Go
Stars: 12,708
Forks: 1,156
Created: 10 Apr 2026
Archived: No

Activity

Last commit: 24 Sep 2026
Latest release: v0.7.2 · 24 Sep 2026
Issues open / closed: 65 / 393 · 86% closed

v0.1.0 (2026-04-20) to v0.7.2 (2026-09-24): a minor roughly every 3–5 weeks, each preceded by 1–3 RCs and followed by patch releases, plus separate guest-kernel releases. Bilingual changelog per version.

Fetched via gh on 25 Sep 2026.

Self-hosting

Methods: binarykuberneteshelmterraformsource-build
You operate: kvmredismysqlobject-storagecontainerdcoredns

Self-host is the only way to run it: free under Apache-2.0 on your own servers. Needs x86_64 or aarch64 Linux with KVM. On ordinary cloud VMs without /dev/kvm, the recommended path installs Tencent's PVM host kernel (x86_64 only). /data/cubelet needs ≥50 GB (200 GB+ recommended). The one-click installer bundles MySQL, Redis, MinIO (S3 for Volumes and snapshots) and CoreDNS (quick-start wildcard DNS only). Production needs your own wildcard DNS and TLS certificates. Helm on Kubernetes and cross-node snapshots are preview. SECURITY: CubeAPI allows every request without authentication until you configure an auth-callback service. The hosted Tencent Cloud Agent Runtime (AGS) is a separate product, not a managed version of this repo.

Documentation signals 8/8

SupportedQuickstart
SupportedAPI reference
SupportedSelf-host guide
SupportedChangelog
SupportedOpenAPI
Supportedllms.txt
SupportedExamples
SupportedArchitecture doc

07 Company & presence Facts only, not scored

Company

Headquarters: Shenzhen, China (Tencent)
Founded: 2023 (internal); open-sourced 2026-04-20
Team size: Not disclosed
Hiring: Unknown

Notable people

Jin Feng: head of the Tencent Cloud IaaS Frontier Technology Team (InfoQ interview, 2026-08-13)

Activity

Last release: v0.7.2 — 2026-09-24 github.com
Last blog post: 2026-09-17 — WeKnora / openFuyao / Huajiao case studies cubesandbox.com

Presence 6

Links found on the public surface scanned; not an exhaustive search.

08 Sources & method How this entry was produced

Public surface only

Everything here was read from cubesandbox.com's unauthenticated pages, docs, pricing and repository on 25 Sep 2026. Nothing behind a login was used.

Statuses, not scores

Capabilities are pass, partial or fail against a written rubric (features v19). “Not assessed” means the scan predates the key.

Confidence is separate

Each fact carries its own confidence. Low-confidence facts are shown, never hidden, so you can weigh them.

Schema: scan v5
Features registry: v19
Keys tracked today: 61
Assessed on this entry: 57

Something wrong on this page?

Nearby in the atlas

All Sandbox products →