AIAny
AI Infra2024
Icon for item

microsandbox

Runs untrusted workloads inside hardware-isolated local microVMs with Docker-like workflows; cross-platform and OCI-compatible, embeddable via SDKs and optimized for fast startups (typical guest boot <100 ms).

Introduction

Why this matters

Hardware-level isolation is a pragmatic way to run code you don't fully trust while minimizing blast radius. As agents, CI jobs, plugin systems, and other automation increasingly execute third-party code, a lightweight microVM runtime that boots quickly and can be embedded into apps lets you keep untrusted work visibly separate from host processes and credentials.

What Sets It Apart
  • Hardware isolation, not just namespaces: uses microVMs with a guest kernel so the security boundary is virtualization hardware rather than process/container namespaces — this reduces many host-escape classes common in container-only sandboxes.
  • Instant, local-first workflow: boots guests in tens of milliseconds (average guest boot under ~100 ms on measured hardware), supports standard OCI images, and exposes Docker-like commands and lifecycle operations so teams reuse existing image workflows.
  • Embeddable SDK + CLI parity: SDKs for Rust, TypeScript, Python, and Go let applications spawn sandboxes as child processes (no long-running daemon). The msb CLI mirrors the same lifecycle, image, and volume commands for interactive and automated use.
  • Agent and secret integration: provides Agent Skills and an MCP server so LLM/agent toolchains can create and control sandboxes; supports host-side secrets with policies that prevent secret material from entering the guest.
Who It's For & Trade-offs

Great fit if you need to run untrusted or third-party workloads with a stronger isolation boundary than containers — examples include AI coding agents, user-submitted code execution, CI job isolation, browser automation, and document processing. The SDK/CLI model makes it convenient to integrate directly into developer tooling or agent workflows.

Look elsewhere if you require a fully managed multi-tenant cloud service today (cloud offering is private beta), or if your target hosts cannot enable required virtualization primitives: macOS requires Apple Silicon, Linux requires KVM, and Windows requires WHP. Also note the project is labeled beta — expect API changes and rough edges while the project matures.

Where It Fits

Positioned between heavyweight VM infrastructure and lightweight containers: it aims to combine the usability and image compatibility of container workflows with the stronger security properties of a microVM. Useful as a drop-in isolation layer for agent-driven automation, CI runners, and developer sandboxes where credential and network egress control matter.

Information

  • Websitegithub.com
  • OrganizationsSuperrad Company, Y Combinator
  • Published date2024/10/03

Categories

More Items

Enables interactive serving of large Mixture-of-Experts (MoE) models on personal machines by adapting offload and execution to measured device bandwidth and agentic workload patterns. Key features include bandwidth-adaptive execution, semantic-aware caching of recurrent state, and an elastic GPU expert cache; supports 20+ MoE models and runs models from ~35B to 753B on consumer/workstation GPUs.

GitHub

Unifies enterprise knowledge into a permission-aware context layer that delivers citation-backed, explainable search and no-code or SDK-driven agentic workflow automation. Supports 50+ connectors, knowledge-graph retrieval, an MCP server, and bring-your-own-model self-hosting.

GitHub
AI Infra2024

Provides a GenLayer DApp boilerplate implementing a football-bets intelligent contract with web and LLM integration. Bundles fast in-memory direct tests with web/LLM mocking, a GenVM linter, integration test workflows, CI, and a Next.js 15 TypeScript frontend.