Blitz Strike
MCP ServerA penetration-testing methodology delivered as one MCP server: BLITZ maps the attack surface, EAGLE-EYE traces source-to-sink, STRIKE live-verifies every finding before it is reported.
- mcp-server
- security
- pentest
- methodology
Give your AI agent a brand-new Windows PC, then throw it away: clean Windows Sandbox guests driven through the sbx CLI and an optional MCP adapter to deploy, drive the UI, catch crashes, keep evidence and delete the machine (Windows Pro/Enterprise/Education, MIT)
sandbox-mcp turns Windows Sandbox into the agent's brand-new customer PC: every run boots a real, disposable Windows 11 desktop — deploy your app, drive its controls over UI Automation, catch the crash, keep the report and audit log, then delete the machine (400 stars as of writing, 2026-11-11). The sbx CLI works standalone; the MCP adapter is an optional second interface over the same engine, and measured GUI samples take roughly 40-50 seconds end to end. Typical jobs: clean-machine smoke tests (what breaks without the VC++ runtime, .NET or WebView2), installer/uninstaller audits (sbx install-test records every file, registry key, service and scheduled task), offline detonation of untrusted files (sbx detonate reports processes, DNS lookups and planted autoruns), and the build-test-on-a-clean-machine-look-fix loop for agents. The default never installs dependencies; explicitly requested provisioning verifies SHA-512 hashes and Microsoft signatures and marks its passes PROVISIONED (not a clean-machine test). The README ships a measured compatibility matrix across seven configurations, and the project site is at https://shanselman.github.io/sandbox-mcp/.
For Windows developers and testers who need the answer to 'does it install, launch and work on a clean customer PC': dependency smoke tests, installer audits, offline detonation of untrusted files, and agent-driven fix loops that never touch your real desktop; also for observing untrusted files inside an isolated, network-off sandbox.
# download the Windows x64 portable ZIP for v1.2.0 from https://github.com/shanselman/sandbox-mcp/releases/tag/v1.2.0 (verify the adjacent .sha256); the ZIP bundles Node.js — no Node/npm install needed
# readiness check (never enables features, elevates, changes PATH or restarts by itself):
.\sbx.cmd doctor
# if the Sandbox feature is off — admin PowerShell, then restart Windows:
Enable-WindowsOptionalFeature -Online -FeatureName Containers-DisposableClientVM -All
# after restarting, open Windows Sandbox from the Start menu once (this installs the wsb CLI that sbx uses)
# optional MCP adapter — the VS Code .vscode\mcp.json shape (replace the example folder with where you extracted the ZIP):
{
"servers": {
"windows-sandbox": {
"type": "stdio",
"command": "C:\\Users\\you\\Downloads\\sandbox-mcp\\node.exe",
"args": ["C:\\Users\\you\\Downloads\\sandbox-mcp\\src\\server.js"]
}
}
}
# for Claude Desktop use "mcpServers" with the same command/args and omit "type" and "tools" (per the README)
# then try it: .\sbx.cmd run C:\Windows\System32\charmap.exe --openWindows-only and edition-gated: Windows Sandbox needs Windows 11 24H2 (build 26100) or later on Pro, Enterprise or Education — Windows Home and Windows Server cannot run it; virtualization must be on in UEFI/BIOS (Intel VT-x / AMD-V); minimum 4 GB RAM (8 GB recommended), 2 CPU cores and 1 GB free disk. One sandbox at a time, and nothing persists by design — output the guest must keep has to be explicitly shared to the host. The portable ZIP is built and tested on native Windows x64; ARM64 is not currently packaged. The default never installs dependencies: provisioning happens only on explicit request, and a provisioned pass is marked PROVISIONED, never a clean-machine pass; the README is explicit that this is not a promise every Windows app or hardware-dependent workload works in Sandbox, and the measured compatibility matrix covers seven configurations. Running inside a VM requires nested virtualization on the host first. MIT. 400 stars as of writing (2026-11-11).
Repository
shanselman/sandbox-mcpCharted
A penetration-testing methodology delivered as one MCP server: BLITZ maps the attack surface, EAGLE-EYE traces source-to-sink, STRIKE live-verifies every finding before it is reported.
Red Hat's "ripgrep for AI context" — a C++23 zero-dependency CLI that also speaks MCP.
One shared session memory for 25+ coding agents — Claude Code, Codex, and Cursor finally remember the same project.