AI Decision System
Replayable, offline-verifiable records for AI-assisted go/no-go decisions: experiment promotions, typed model answers and human-approved actions.
AI 기반 Windows 시스템 관리, 파일 복구·보안·진단 지원
Windows system administration through AI with file recovery, security, and diagnostics
개발 도구Python · TypeScriptGitHub @sandraschi직접 설치운영자 미확인
써 보셨어요? 한 번만 눌러 주세요. 로그인은 필요 없어요.
아직 감지된 업데이트가 없습니다. GitHub 릴리스가 나오면 여기에 보입니다.
개인 계정의 공개 저장소 소유자는 GitHub로 로그인해 운영자를 확인하고 소개를 직접 관리할 수 있습니다. 관리자 승인은 필요하지 않습니다.
로그인한 뒤 확인 파일을 만들어 이 저장소의 기본 브랜치에 30분 안에 커밋해 주세요. 다른 브랜치나 포크의 파일은 인정하지 않습니다. 파일을 올린 뒤 아래에서 확인하면 관리자 승인 없이 관리할 수 있습니다. 자격을 유지하려면 파일을 남겨 두세요.
로그인이 필요하면 로그인 화면으로 이동합니다. 로그인 후 이곳으로 돌아와 원하는 확인 버튼을 다시 눌러 주세요.
설치형 프로젝트는 배포 도메인 없이도 확인할 수 있습니다. 개인 저장소는 GitHub 로그인으로, 조직 저장소는 위의 저장소 확인 파일로 직접 확인해 주세요.
README 전체를 불러오는 중입니다. 아래는 저장된 발췌입니다.
📖 Installation Guide — quick start, manual setup, and troubleshooting Windows system administration, through AI. File recovery, security, disk maintenance, diagnostics, services, processes — all accessible via MCP (Claude Desktop, Cursor, etc.) and a React web dashboard. ⚠️ Administrator privileges required. Disk operations, service management, file recovery, and permission changes need elevation. Run your terminal as Administrator before starting the server.README 전문 보기 ↗
📖 Installation Guide — quick start, manual setup, and troubleshooting
Windows system administration, through AI. File recovery, security, disk maintenance, diagnostics, services, processes — all accessible via MCP (Claude Desktop, Cursor, etc.) and a React web dashboard.
Download the installer: latest release (
system-admin-mcp-0.4.0-setup.exe) — no Python needed, backend embedded. For Claude Desktop without installing: attachsystem-admin-mcp-0.4.0.mcpbfrom the same release.Claude Desktop only — one-line install of the latest bundle:
irm https://github.com/sandraschi/system-admin-mcp/releases/latest/download/install.ps1 | iex
⚠️ Administrator privileges required. Disk operations, service management, file recovery, and permission changes need elevation. Run your terminal as Administrator before starting the server.
// Claude Desktop — from source (no registry package; see INSTALL.md):
{ "mcpServers": { "system-admin-mcp": {
"command": "uv",
"args": ["--directory", "D:\\Dev\\repos\\system-admin-mcp", "run", "system-admin-mcp"]
} } }
git clone https://github.com/sandraschi/system-admin-mcp
cd system-admin-mcp
just
This opens an interactive dashboard showing all available commands. Run just bootstrap to install dependencies, then just serve or just dev to start. Frontend deps install via Bun 1.3+ (bun.lock committed); Node 22+ stays for the Vite/Tauri CLIs.
If you don't have just installed:
git clone https://github.com/sandraschi/system-admin-mcp
cd system-admin-mcp
uv sync --all-extras
uv run system-admin-mcp # stdio mode (for MCP clients like Claude Desktop)
uv run system-admin-mcp --web # HTTP mode on :10861 (web dashboard backend)
Verify: Invoke-WebRequest http://127.0.0.1:10861/api/health -UseBasicParsing → 200.
See INSTALL.md for prerequisites, installer/MCPB options, and troubleshooting.
| Area | Operations |
|---|---|
| System Health | CPU/RAM/disk metrics, event logs, hardware inventory, installed software |
| File Recovery | Scan NTFS MFT, recover deleted files, validate integrity |
| Security | View/set/audit NTFS permissions, take ownership, network port audit, Defender/VPN/Tailscale posture, red-button Airgap |
| Disks | SMART health, defrag (HDD), TRIM (SSD), cleanup, folder size analysis |
| Services | List/start/stop, change startup type, paginated |
| Processes | List/sort/analyze/kill, paginated, sortable by CPU/Memory/Name/PID |
| Startup & Taskbar | Manage startup programs, toggle autohide, find blockers, taskbar buttons + tray icons with autostart flags |
| AI Chat | Local (Ollama/LM Studio) + 13 cloud providers via backend proxy, streaming, resident-first GPU placement |
| Safety | Read-only mode (SYSTEMADMIN_READ_ONLY=1), mutation audit log, confirm-gated Airgap (see docs/SECURITY.md) |
| Agentic | Let AI autonomously diagnose and fix issues (SEP-1577 sampling) |
All operations go through a single system_admin tool — one tool, 40+ operations.
| Feature | What it provides |
|---|---|
ctx.sample() (SEP-1577) | Server-side LLM calls for autonomous diagnosis. Tools agentic_system_workflow and autonomous_system_troubleshooter use ctx.sample() to orchestrate multi-step diagnostics without client round-trips — the server borrows the client LLM to interpret system data and recommend actions. |
| Prompts (4 templates) | Registered via @mcp.prompt(): system_diagnostics_expert, security_hardening_expert, system_troubleshooter, volume_maintenance_expert. Each has parameterized focus modes (e.g. performance vs events, ownership vs audit). Clients inject them as system instructions. |
Skills (skill://) | SkillsDirectoryProvider exposes skill://system-admin-expert/SKILL.md — a portable expertise document with safe-mode patterns, diagnostic sequences, and SEP-1577 workflow recipes. Discoverable via resources/list. |
| Prefab UI (4 cards) | @mcp.tool(app=True) tools returning ToolResult with structured_content=PrefabApp(...): system_health_card (CPU/RAM/disk), top_processes_card (sorted by CPU/memory), list_services_card (filtered services), volume_status_card (all volumes with bar charts). Rendered natively by capable hosts (Claude Desktop side-panel). |
| CodeMode | BM25 discovery for agentic tool selection. Enabled via --agentic flag. |
| SkillsDirectoryProvider | Registers skill:// resources from skills/ directory. |
| Dual transport | stdio (default for Claude Desktop) + streamable-http (MCP_TRANSPORT=http). |
A 20-page React SPA on ports 10860 (frontend) / 10861 (backend):
just web # Backend API
just web-frontend # Frontend dev server
Pages: Dashboard, Inbox (attention feed), Status, Processes (paginated, sortable), Services (paginated), Taskbar (buttons + tray + autostart), Volumes, File Owner, File Recovery, Logs, Tools, Apps, Elevated, Chat, Settings, Help.
| Command | What it does |
|---|---|
just dev | Start MCP server (Claude Desktop) |
just serve | Start MCP via HTTP |
just test | Run tests |
just lint | Check code quality (ruff + biome) |
just fix | Auto-fix everything |
just certify | Full gates: ruff + format + pyright + pytest + frontend build + biome |
just mcpb-pack | Build MCPB bundle |
just build | Install dependencies |
just web | Start FastAPI backend on 10861 |
just web-frontend | Start Vite frontend on 10860 |
React 19 + Vite 7 + TailwindCSS 3 + Radix UI + Lucide + TanStack Query +
Zustand 5 (LLM/GPU store) + React Router 7. Backend: FastAPI + FastMCP 3.4
(fastmcp[tasks]>=3.4.4,<4) on Python 3.12+, uv managed. No Bootstrap,
no jQuery, no Redux.
| Var | Default | Effect |
|---|---|---|
SYSTEMADMIN_READ_ONLY=1 | off | Mutation kill switch |
SYSTEMADMIN_LLM_KEYSTORE | data/llm_keys.json | 0600 LLM key store |
MCP_TRANSPORT=http | stdio | Streamable-HTTP transport |
WEBAPP_PORT / PORT | 10861 | Backend port |
Full install + troubleshooting: INSTALL.md.
├── AGENTS.md # Instructions for AI agents
├── justfile # Task runner (SOTA Industrial Dashboard)
├── pyproject.toml # Python project config (FastMCP 3.4)
├── src/system_admin_mcp/
│ ├── app.py # FastMCP instance + lifespan (skills, prefabs)
│ ├── server.py # FastAPI backend (20+ REST endpoints)
│ ├── transport.py # stdio / streamable-http dual transport
│ ├── prompts.py # 4 @mcp.prompt() templates
│ ├── tools/
│ │ ├── portmanteau.py # system_admin (40+ ops dispatcher)
│ │ ├── implementations.py # File recovery, ACLs, disk, WMI
│ │ ├── services_and_tasks.py # Services, processes, startup, taskbar
│ │ ├── agentic_system_workflow.py # ctx.sample() workflows
│ │ ├── monitoring.py # Watchdog file watcher
│ │ ├── system_ops.py # Standalone tools
│ │ └── prefab/ # 4 Prefab UI cards
│ ├── elevated_service/ # Named-pipe elevated bridge
│ └── user_bridge/ # Service communication client
├── web_sota/ # React 19 + Vite dashboard
├── mcpb/ # MCPB packaging (v0.2 manifest)
├── skills/ # skill://system-admin-expert/SKILL.md
└── tests/ # Pytest suite
| Document | For |
|---|---|
| Quickstart | End users — install, configure, run |
| API Reference | Developers — REST endpoints |
| Development | Contributors — hacking on the server |
| AGENTS.md | AI agents — architecture & patterns |
Built by Sandra Schipal in Vienna. MIT.