convez.ai
MCP ServerDevOps & InfraestruturaAvançado
clayboicardi

MSI Center MCP Server

clayboicardi/msi-center-mcp

Decodifica MSI Center e valida prontidão da máquina para tarefas de IA

Ver no GitHub
InicialPluginScore · de 0 a 100

Servidor MCP que oferece aos agentes de IA visibilidade completa sobre máquinas MSI Windows, decodificando configurações do MSI Center diretamente do registro, telemetria de hardware, saúde da bateria e verificações de prontidão por tarefa. Permite que Claude e outros clientes MCP compreendam e validem se a máquina está configurada corretamente para workloads específicos sem fazer alterações.

Instalação

$npx convez-ai clayboicardi/msi-center-mcp

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
57%
Último commit
há 2 meses
Versão
—
Issues abertas
1
Contribuidores
1

Casos de uso

  • Validar configuração da máquina antes de treinar modelos de IA
  • Diagnosticar e explicar configurações opacas do MSI Center
  • Monitorar telemetria em tempo real (temperatura, potência, RAM, VRAM)
  • Automatizar verificações de pré-requisitos de hardware para workloads específicos

Veredito

Ferramenta altamente especializada para o nicho de proprietários de laptops MSI que executam IA localmente. Código bem documentado, honesto sobre limitações (confiança por máquina) e sem pretensões — resolve um problema real de opacidade do MSI Center. Requer calibração manual para máquinas não documentadas, o que é transparente e até bem-vindo.

README

A tradução para o português deste README ainda está na fila.

msi-center-mcp

A read-only MCP server that gives AI coding agents (Claude Code, or any MCP client) full insight into an MSI Windows machine: hardware telemetry, Windows power state, live MSI Center settings decoded from the registry, battery health, a curated knowledge base of what each MSI setting actually does, and per-task readiness checks — e.g. "is this machine configured correctly to train a model right now?"

> check_profile_readiness {"profile": "llm_training"}
ready: true — 8/8 rules pass
  ✓ AC power            ✓ GPU present (nvidia-smi)
  ✓ MSI User Scenario = Extreme Performance   ✓ Windows plan = High performance
  ✓ GPU cool (43°C)     ✓ VRAM free            ✓ RAM headroom   ✓ AI Engine off

Why

MSI Center is opaque: its settings live in an Embedded Controller behind an undocumented UI, and nothing on the Windows side documents what's actually configured. It turns out MSI Center's service mirrors its user-facing state to readable registry keys (HKLM\SOFTWARE\WOW6432Node\MSI\MSI Center\...) — no admin required, ~80 ms per read. This server decodes that state, explains it, and checks it against what your workload needs. It changes nothing: when something is misconfigured, it tells you exactly what to click in MSI Center or Windows.

Calibration model (read this before trusting it)

Registry integers mean different things on different MSI machines. Decoded values carry a confidence label:

  • verified_live — observed on a calibrated machine (UI action diffed against the registry)
  • community — documented by community projects (msi-ec, MControlCenter)
  • inferred — plausible hypothesis (UI ordering, single observation)
  • unknown — no mapping; readiness checks report unknown, never a false failure

Machine profiles are selected at runtime from registry signals (EC firmware signature on notebooks; platform type + board model on desktops). Currently calibrated:

  • MSI Vector A16 HX A8WHG notebook (Ryzen 9 8940HX, RTX 5070 Ti Laptop 12GB, MSI Center 2.0.70)
  • MSI PRO B760M-VC WIFI (MS-7D37) desktop (i5-14400F, RTX 4060, MSI Center SDK 3.2026) — desktop MSI Center uses a different registry family entirely (string scenario names under SyncData, per-fan state under Setting\FANn, Zero Frozr)

Everything else gets the generic profile at reduced confidence — still useful, honestly labeled. Calibrating your own machine takes ~10 minutes (see docs/msi-center-knowledge.md); contributions of new machine profiles are welcome.

MCP Tools (15)

CategoryTools
Telemetryget_system_info, get_power_status, list_power_plans, get_active_power_plan, get_gpu_snapshot, get_telemetry_snapshot, capture_telemetry_log, summarize_telemetry_log, compare_telemetry_logs
MSI insightget_msi_center_state, get_battery_health, explain_msi_setting
Profileslist_profiles, dry_run_profile, check_profile_readiness

Profiles: balanced_daily, gaming_ac, quiet_work, battery_saver, cooldown, llm_inference, llm_training.

Typical agent flows:

  • "Are we set up to train?" → check_profile_readiness {"profile": "llm_training"} → follow the manual_fix strings for any failed rule.
  • "What does Cooler Boost actually do?" → explain_msi_setting {"topic": "fan_modes_cooler_boost"}.
  • "Did switching scenarios help?" → capture_telemetry_log before/after → compare_telemetry_logs.

Quick start

Requirements: Windows 11, Node.js 20+. Optional (each degrades gracefully): NVIDIA driver with nvidia-smi on PATH, MSI Center installed.

git clone https://github.com/clayboicardi/msi-center-mcp.git
cd msi-center-mcp
npm install
npm run build

Wire into Claude Code (user scope = available in every session):

claude mcp add --scope user msi-center -- node <absolute-path-to-repo>\dist\index.js

Any other MCP client: stdio transport, command node <repo>\dist\index.js.

What it does NOT do

  • No writes of any kind: no power plan changes, no MSI Center changes, no fan/clock/voltage changes, no registry writes (reg.exe is allowlisted to query on two fixed keys).
  • No arbitrary command or PowerShell execution — every subprocess goes through an allowlisted command runner (execFile, argument arrays, shell: false, timeouts), with safety tests asserting that write-shaped arguments are rejected.
  • No admin elevation (the elevated WMI/EC interface MSI Center itself uses was evaluated and rejected).
  • No network access at runtime.

See SECURITY.md for the full threat model and the checklist any future write feature must pass.

Performance notes

  • The PowerShell CIM query costs ~3-4 s per invocation (powershell.exe spawn). Every snapshot-style tool runs it exactly once per call; get_msi_center_state (~80 ms) and get_battery_health (~2 ms) avoid it entirely.
  • capture_telemetry_log keeps middle samples lean (nvidia-smi + OS + powercfg only) so 1-2 s sampling intervals actually hold; first and last samples are full so battery drain is measurable. Keep durationSeconds under your MCP client's tool timeout (30-60 s is plenty).

Configuration

Zero-config by default. Environment overrides:

VariableDefaultMeaning
MSI_CENTER_MCP_LOGS_DIR<package>/logsTelemetry log directory
MSI_CENTER_MCP_COMMAND_TIMEOUT_MS10000Subprocess timeout
MSI_CENTER_MCP_MAX_LOG_DURATION_SECONDS600Capture duration ceiling
MSI_CENTER_MCP_ALLOW_EXTERNAL_LOG_PATHSfalseAllow log reads outside the logs dir

Development

npm test          # 76 tests, fake providers — no MSI hardware needed
npm run lint
npm run typecheck
node scripts/smoke-real.mjs   # manual real-hardware smoke check (read-only)

Project docs: docs/msi-center-knowledge.md (what the settings do + calibration guide), AGENTS.md (rules for AI agents working on this repo), SECURITY.md. The repo's origin story — v0.1 scaffolded by OpenAI Codex, then reviewed, fixed, extended, and calibrated by Claude Code — is preserved in docs/claude-code-handoff.md (historical).

License

MIT

Ver o original no GitHub

Bom para

Relacionados