← Back to research
·8 min read·company

Microsandbox

Microsandbox is an open-source, self-hosted sandbox platform using libkrun microVMs for hardware-isolated AI agent code execution with sub-200ms startup times.

Key takeaways

  • Local-first microVM sandboxes with hardware isolation and sub-200ms startup — secrets never leave the host
  • libkrun-based architecture provides stronger isolation than containers with near-container startup speed
  • YC X26 batch with 5K+ GitHub stars — early stage but shipping fast with SDKs for Python, JavaScript, and Rust

FAQ

What is Microsandbox?

Microsandbox is an open-source tool for running untrusted code in hardware-isolated microVMs on your own machine, designed for AI agents that need secure code execution without cloud dependencies.

How does Microsandbox differ from E2B?

Microsandbox is self-hosted and local-first using libkrun microVMs, while E2B is a cloud-hosted platform using Firecracker. Microsandbox keeps secrets on-host; E2B offers managed infrastructure at scale.

What languages does Microsandbox support?

Microsandbox provides SDKs for Python, JavaScript/Node, and Rust, with OCI-compatible images supporting any language runtime.

Is Microsandbox production-ready?

Microsandbox is still experimental. It works on macOS and Linux, with Windows support in progress. Expect breaking changes as the project matures.

Executive Summary

Microsandbox is an open-source, self-hosted sandbox platform that uses libkrun microVMs to provide hardware-isolated execution environments for AI agents. Built in Rust by Zerocore AI (YC X26), it offers sub-200ms startup times with true VM-level isolation — combining the security of traditional VMs with the speed of containers.

AttributeValue
CompanyZerocore AI
Founded2024
FundingY Combinator (X26 batch)
FounderStephen Akinyemi
HeadquartersUnknown

Product Overview

Microsandbox takes a fundamentally different approach from cloud-hosted sandbox platforms like E2B. Instead of sending code to remote infrastructure, Microsandbox runs lightweight microVMs locally on your own machine. Secrets never leave the host, networking is programmable from the outside, and each sandbox gets its own dedicated kernel.

The project was created by Stephen Akinyemi, who previously built infrastructure at Zerocore AI. He has articulated the thesis that containers are insufficient for running truly untrusted AI-generated code — shared kernels mean sophisticated attacks can escape. MicroVMs solve this with hardware-level isolation while maintaining the speed developers expect.

Key Capabilities

CapabilityDescription
MicroVM IsolationEach sandbox runs in its own VM with a dedicated kernel via libkrun
Sub-200ms StartupBoot a full VM faster than most container solutions
Secret InjectionCredentials injected at the network layer — the guest never sees real values
Programmable NetworkingInspect DNS, analyze HTTP traffic, block exfiltration at the IP level
OCI CompatibleWorks with standard container images (Docker Hub, GHCR)
MCP ServerBuilt-in Model Context Protocol server for AI agent integration

Product Surfaces / Editions

SurfaceDescriptionAvailability
CLImsb command for server, sandbox, and project managementBeta
Python SDKpip install microsandboxBeta
JavaScript SDKnpm install microsandboxBeta
Rust SDKcargo add microsandboxBeta
MCP ServerNative MCP integration for AI agentsBeta
Project SandboxesSandboxfile-based project config (like package.json for sandboxes)Beta

Technical Architecture

Microsandbox uses libkrun, a lightweight virtualization library from the containers project that leverages KVM (Linux) or Hypervisor.framework (macOS) for hardware isolation. Unlike Firecracker (used by E2B and AWS Lambda), libkrun includes its own Virtual Machine Monitor and uses Transparent Socket Impersonation (TSI) for networking — avoiding the overhead of virtual network interfaces.

┌──────────────────────────────────────┐
│          Host Machine                │
├──────────────────────────────────────┤
│  ┌──────────┐  ┌──────────┐          │
│  │ Sandbox  │  │ Sandbox  │   ...    │
│  │ (microVM)│  │ (microVM)│          │
│  │ own      │  │ own      │          │
│  │ kernel   │  │ kernel   │          │
│  └────┬─────┘  └────┬─────┘          │
│       │              │               │
│  ┌────┴──────────────┴─────┐         │
│  │   libkrun (KVM / HVF)   │         │
│  └─────────────────────────┘         │
│  ┌─────────────────────────┐         │
│  │   msb server (Rust)     │         │
│  │   + secret injection    │         │
│  │   + network policy      │         │
│  └─────────────────────────┘         │
└──────────────────────────────────────┘

Key Technical Details

AspectDetail
Isolationlibkrun microVMs (hardware-level, KVM/HVF)
Cold Start~187ms (measured)
PersistenceFiles saved to ./menv directory on host
LanguageRust (core), with Python/JS/Rust SDKs
Open SourceYes (Apache 2.0)
Self-HostingSelf-hosted only — no cloud service (yet)
PlatformsmacOS (working), Linux (working), Windows (WIP)

Secret Management — The Killer Feature

Microsandbox's most innovative feature is network-layer secret injection. When you provide an API key with an allowed host:

  1. The sandbox sees only a random placeholder (e.g., msb_placeholder_a7f3...)
  2. When the sandbox makes a verified TLS connection to the allowed host, Microsandbox swaps the placeholder for the real credential at the network layer
  3. If the sandbox tries to send the placeholder to any other host, only the worthless placeholder is transmitted

This means even if malicious code reads environment variables and tries to exfiltrate them, the real credentials never leave the host. DNS rebinding protection, cloud metadata blocking, and DNS-to-IP binding activate automatically.


Strengths

  • True hardware isolation — Each sandbox has its own kernel; container escape attacks are irrelevant
  • Local-first — Secrets and data never leave your machine; no cloud dependency
  • Secret injection — Network-layer credential management prevents exfiltration by design
  • Fast startup — Sub-200ms boot for a full microVM is impressive, nearly matching container speed
  • Programmable networking — DNS inspection, HTTP interception, DLP, and domain allowlisting built in
  • Apache 2.0 — Permissive license with no AGPL restrictions (unlike Daytona)
  • Rust codebase — Memory-safe foundation for security-critical infrastructure
  • OCI compatible — Use standard Docker/container images; no proprietary format

Cautions

  • Experimental — Project explicitly warns of breaking changes and missing features
  • No cloud offering — Self-hosted only; teams wanting managed infrastructure need to look elsewhere
  • Small team — Community-driven with ~3 contributors; bus factor is a concern
  • No GPU support — CPU-only microVMs; ML inference workloads need alternatives
  • Windows not ready — Only macOS and Linux are working; Windows is WIP
  • Limited ecosystem — 5K stars vs E2B's 6K+ with 1M monthly SDK downloads; early adoption curve
  • No enterprise features — No SOC2, no RBAC, no audit logging (yet)

Pricing and Licensing

AspectDetail
LicenseApache 2.0 (permissive)
CostFree and open source
Cloud Service"Launching soon" per website
Self-HostedFully supported, single curl install

Microsandbox is currently free. The website indicates a cloud product is coming ("Launching soon"), which will presumably be the monetization path. YC backing suggests a cloud-hosted or managed version is on the roadmap.


Competitive Positioning

Direct Competitors

CompetitorDifferentiation
E2BE2B is cloud-hosted with Firecracker; Microsandbox is local-first with libkrun. E2B has enterprise scale; Microsandbox has stronger secret management
DaytonaDaytona is a full dev platform (Docker-based); Microsandbox is a focused execution primitive with hardware isolation
ModalModal offers serverless GPU compute; Microsandbox is local, CPU-only, security-focused
CodeSandboxCodeSandbox targets browser-based dev; Microsandbox targets agent-driven code execution

When to Choose Microsandbox Over Alternatives

  • Choose Microsandbox when: Running untrusted AI-generated code locally, secrets must stay on-host, or you need maximum isolation without cloud costs
  • Choose E2B when: You need managed infrastructure at enterprise scale with compliance certifications
  • Choose Daytona when: You need a complete development platform with team collaboration features
  • Choose Modal when: You need GPU access for ML workloads

Ideal Customer Profile

Best fit:

  • AI agent developers who need to execute untrusted code locally with maximum security
  • Teams handling sensitive API keys who cannot tolerate credential exposure to third parties
  • Self-hosting enthusiasts who want cloud-grade isolation on their own hardware
  • Developers building MCP-integrated tools that need sandboxed execution
  • Security-conscious startups building coding agents or data pipelines

Poor fit:

  • Enterprise teams needing managed cloud infrastructure with SLAs
  • Organizations requiring SOC2 compliance and audit trails
  • Teams needing GPU access for ML workloads
  • Windows-only development environments
  • Teams wanting a mature, battle-tested production platform today

Viability Assessment

FactorAssessment
Financial HealthEarly — YC X26 backing, no announced funding round
Market PositionChallenger — Differentiated on local-first and secret management
Innovation PaceRapid — Active development, frequent commits, new features shipping
Community/EcosystemGrowing — 5K+ GitHub stars, 238 forks, active Discord
Long-term OutlookPromising — Strong technical thesis, YC backing, but needs to ship cloud product and grow team

Microsandbox occupies a unique position in the AI sandbox space: it is the only major player betting entirely on local-first, self-hosted microVM isolation. If the thesis is correct that AI agents will increasingly run on developer machines (not just in the cloud), Microsandbox could become the default execution layer. The secret injection feature alone is a meaningful innovation that competitors will likely copy.


Bottom Line

Microsandbox is the most security-focused AI sandbox available today. The combination of hardware-isolated microVMs, network-layer secret injection, and programmable networking makes it technically superior to container-based alternatives for running truly untrusted code.

The trade-off is maturity. This is experimental software from a small team with no cloud offering yet. You are betting on the project and the thesis, not on proven enterprise infrastructure.

Recommended for: Developers building AI agents locally who need maximum isolation and cannot tolerate secret exposure to third-party infrastructure.

Not recommended for: Enterprise teams needing managed, compliant, production-grade sandbox infrastructure today.

Outlook: With YC backing and a compelling technical story, expect a cloud product launch in 2026. The local-first positioning and secret management innovation give Microsandbox a real shot at carving out a significant niche alongside E2B. Watch for the cloud launch and first enterprise customers.


Research by Ry Walker Research • methodology