Blaxel Sandbox

license

MIT(OSI)

source

public(sandbox API, self-hostable via Docker Compose)

managed-tier

Blaxel proprietary perpetual-standby platform

core-gated

gated(the control plane that creates, suspends and resumes sandboxes - the perpetual-standby platform Blaxel sells - exists in no public repo

MIT sandbox API and image hub with a proprietary managed platform. The core is gated on a named mechanism rather than merely because a paid tier exists, and nothing is carved out inside the repository: the piece that makes Blaxel a sandbox product is in no repository at all. blaxel-ai/sandbox is 419 entries of sandbox-api - the HTTP API running inside a sandbox - plus about thirty Dockerfile templates under hub/, and the README calls it "Sandbox Hub ... a collection of development environment templates". The scheduler that allocates a micro VM, and the perpetual standby and sub-25ms resume the product is sold on, appear nowhere across the 29 public repos in the org, which hold SDKs, plugins and forks of third-party infrastructure. That is decided on what this org publishes rather than by analogy with its neighbors: E2B publishes the infrastructure powering its cloud and beta9 publishes its whole gateway, and both come out fully open, while this product and Tensorlake do not.

• https://github.com/blaxel-ai/sandbox recorded 2026-07-09

  MIT license; sandbox API self-hostable via Docker Compose (port 8080)

• https://blaxel.ai/sandbox recorded 2026-07-09

  perpetual sandbox platform for AI agents; standby at zero compute cost, sub-25ms resume

• https://api.github.com/repos/blaxel-ai/sandbox/git/trees/main?recursive=1 recorded 2026-08-12

  Full untruncated recursive tree, 419 entries, one MIT LICENSE. The entire tree is sandbox-api/ (the HTTP API that runs inside a sandbox: filesystem, process, network), hub/ (about thirty Dockerfile templates - base-image, py-app, ts-app, chromium, jupyter, claude-code), docker/, e2e/ and a docker-compose.yaml for building and testing those images. There is no scheduler, allocator, control plane or standby/resume implementation anywhere in it.

• https://raw.githubusercontent.com/blaxel-ai/sandbox/main/README.md recorded 2026-08-12

  The repo describes itself as "Sandbox Hub ... a collection of development environment templates for creating secure, isolated micro VM environments", and its prerequisites and installation steps are Docker Compose for building and running those template images locally. It documents no way to run the Blaxel platform itself.

• https://api.github.com/orgs/blaxel-ai/repos?per_page=100 recorded 2026-08-12

  All 29 public repos in the org. They are SDKs and clients (sdk-python, sdk-typescript, sdk-go, toolkit, blaxel-mcp-server), integration plugins, image bases and forks of third-party infrastructure (virtual-kubelet, fluent-bit, kubernetes-event-exporter). No repo implements the sandbox control plane the SDKs call; bl-controller is a Kubernetes controller for legacy Beamlit resources.

• https://api.github.com/repos/blaxel-ai/sandbox recorded 2026-08-12

  Repo metadata: language Go, license spdx_id MIT, 26 stars, default branch main.

Very nascent: 27 stars and 13 forks on the sandbox repo, the only declared repo, which falls under 1K stars. With a low-double-digit star base and no other instrument, that reads at level 1. No usage or download signal exists: blaxel.ai/sandbox publishes no customer, request or sandbox-count figure at all - it sells on 25ms resumes and perpetual standby - so stars are all there is. The company is backed by a $7.3M seed (First Round, YC S2025). It is included here on openness and mission-fit grounds despite the low prominence, and scored on what the evidence supports. Confidence is low, and a star count moves, so the figure below is a snapshot.

• https://api.github.com/repos/blaxel-ai/sandbox recorded 2026-08-13

  Repository metadata - stargazers_count 27 and forks_count 13 for blaxel-ai/sandbox, language Go, license spdx_id MIT, public and not archived

• https://blaxel.ai/sandbox recorded 2026-08-13

  No usage, customer or request figure is published anywhere on the product page; it sells on 25ms resumes, perpetual standby and auto-suspend after 15s of inactivity

MicroVM isolation and the perpetual-standby resume model are a genuinely strong, distinctive feature set, but proven scale is unverified and the product is early, so held at 3 rather than level with the proven microVM 4s. The vendor page carries microVMs, ~25ms resume from standby, full filesystem and memory snapshots, auto-suspend after 15s idle, persistence across months, batch jobs across thousands of sandboxes, outbound allow-lists, static IPs, and self-hosted and VPC deployment options. Independent ComputeSDK TTI benchmarks put Blaxel at a 0.51s median time-to-interactive, 0.60s P95, composite 94.5, 9th of 24 sandbox providers - mid-pack, sub-second, and ahead of E2B, Modal, Vercel and Cloud Run. Latency is not what holds the band at 3; the missing evidence is proven scale.

• https://blaxel.ai/sandbox recorded 2026-08-13

  MicroVMs that boot in milliseconds and resume in ~25ms from standby with full filesystem and memory snapshots; sandboxes auto-suspend after 15s of inactivity and persist across sessions; batch jobs spawn thousands of tasks; outbound allow-lists, static IPs and proxy routing; managed cloud, self-hosting and VPC deployment options

• https://www.computesdk.com/benchmarks/sandboxes/ recorded 2026-08-13

  ComputeSDK TTI leaderboard, independent, 100 iterations per run, last run 7 August 2026, 24 providers on 4 vCPU / 16GB in Northern Virginia. Blaxel records 0.51s median, 0.60s P95, 0.61s P99, 100% success, composite 94.5, ranking 9th of 24 by composite score.