Skip to main content
Glama
deenesjakoruzh
LukeLamb

claude-rocm-mcp

by LukeLamb
OverviewSchemaSourceRelated ServersDiscussionsAdmin
JavaScript
Local

Monitor de GPU Claude ROCm

Monitorización de solo lectura de GPU AMD para Claude Desktop en Linux. Permite que Claude responda preguntas como "¿mi entrenamiento es estable?", "¿qué temperatura tiene la GPU?", "¿cuánta VRAM queda?" y "¿qué proceso está usando la tarjeta?", todo ello sin la capacidad de finalizar procesos, cambiar frecuencias de reloj o alterar de cualquier otra forma una carga de trabajo en ejecución.

Utiliza rocm-smi (y opcionalmente amdgpu_top) y expone cinco herramientas estrictamente de solo lectura. Sin finalización de procesos, sin cambios de reloj/energía, sin control de ventiladores. Para esas funciones, utiliza una herramienta como gpu-kill.

Requisitos

  • Linux con una GPU AMD (tarjetas de consumo RDNA2/3/4 — RX 6000/7000/9000, R9700 — y tarjetas de centro de datos CDNA).

  • Módulo del kernel amdgpu cargado (estándar en Ubuntu/Debian/Fedora modernos).

  • rocm-smi instalado:

    sudo apt install rocm-smi            # minimal
# or a full ROCm stack:
# https://rocm.docs.amd.com/projects/install-on-linux/

Algunas métricas (relojes, energía, temperaturas detalladas) requieren versiones más recientes de ROCm. Si una métrica no es compatible con tu tarjeta/controlador, la herramienta la devuelve como null en lugar de fallar.

Related MCP server: System Monitor MCP Server

Instalación (Claude Desktop)

  1. Descarga el archivo ROCm.mcpb más reciente desde la página de Releases.

  2. En Claude Desktop, abre Settings → Extensions.

  3. Desplázate hasta Extension Developer en la parte inferior, haz clic en Install Extension y selecciona el archivo .mcpb.

  4. Habilita la extensión. Las herramientas aparecerán con el prefijo rocm-gpu-monitor: en el selector de herramientas de Claude.

Herramientas

Herramienta

Qué hace

gpu_status

Resumen rápido: nombre, % de utilización, VRAM usada/total, temperaturas (borde/unión/memoria), energía promedio/máxima, %/RPM del ventilador.

gpu_metrics

Volcado completo de rocm-smi -a --json: relojes, voltajes, enlace PCIe, versiones de firmware, actividad por motor.

gpu_processes

Procesos de computación que usan la GPU (PIDs KFD) con uso de VRAM e índice de tarjeta.

gpu_watch

Toma N muestras en un intervalo fijo y devuelve los fotogramas sin procesar más estadísticas de mín/máx/promedio por tarjeta.

rocm_info

Paquetes ROCm/HIP instalados, versión del controlador, estado de carga del módulo amdgpu, disponibilidad de amdgpu_top.

Todas las herramientas incluyen las anotaciones de MCP readOnlyHint: true / destructiveHint: false / openWorldHint: false.

Ejemplos de prompts

"¿Es estable mi entrenamiento? Monitoriza la GPU durante 20 segundos."

"¿Cuánta VRAM está usando mi proceso actual de PyTorch?"

"¿La GPU está sufriendo estrangulamiento térmico? ¿Cuál es la temperatura de unión?"

"¿Qué versión de ROCm está instalada y está cargado el controlador?"

Política de privacidad

Esta extensión se ejecuta completamente en tu máquina local y solo interactúa con los siguientes programas, siempre en modo de solo lectura:

  • rocm-smi — consulta al controlador del kernel AMD para obtener métricas del dispositivo.

  • amdgpu_top (si está instalado) — métricas adicionales más detalladas.

  • dpkg -l — enumera los paquetes relacionados con ROCm instalados por nombre/versión.

  • lsmod — comprueba si el módulo del kernel amdgpu está cargado.

  • which — localiza los binarios anteriores al inicio.

No sale ningún dato de tu máquina. Esta extensión no realiza E/S de red, no abre sockets y no escribe archivos fuera de la salida estándar stdout/stderr del proceso (capturada por el directorio de registros de Claude Desktop). No puede modificar el estado de la GPU (relojes, energía, ventilador), finalizar procesos ni alterar de otro modo una carga de trabajo en ejecución; cada herramienta es estrictamente de solo lectura.

La información visible para Claude incluye el modelo de GPU y las versiones de firmware, el uso actual de VRAM, lecturas de temperatura/reloj/energía, los PIDs de los procesos que utilizan la GPU y su uso de VRAM, y una lista de paquetes instalados relacionados con ROCm. Si consideras que algo de esto es confidencial (por ejemplo, la lista de PIDs de procesos en ejecución en una máquina compartida), no habilites esta extensión.

Solución de problemas

"rocm-smi is not installed" — instálalo (sudo apt install rocm-smi). Si usas una distribución que no es Debian, sigue las instrucciones de instalación de ROCm de AMD.

Muchos campos devuelven null en una GPU nueva — algunas métricas (relojes, energía, temperatura) dependen del soporte del firmware expuesto a través de rocm-smi. En tarjetas muy nuevas (p. ej., RDNA4/R9700), es posible que rocm-smi --showclocks/--showpower informen "No JSON data to report" aunque -a --json devuelva los datos. Esta extensión utiliza -a --json como fuente principal para maximizar la cobertura. Si un campo específico no es compatible con tu tarjeta, será null en lugar de un error.

Errores de permisos — rocm-smi normalmente se ejecuta sin privilegios de root. Si ves errores de permisos, comprueba que el usuario que ejecuta Claude Desktop esté en los grupos render y video:

sudo usermod -aG render,video $USER
# log out and back in

Desarrollo

El servidor es un único script de Node.js de unas 300 líneas sin dependencias de npm. Para reconstruir el .mcpb:

cd bundle-source
zip -j ../ROCm.mcpb manifest.json package.json server.js README.md LICENSE icon.png

Licencia

MIT. Consulta LICENSE.

Relacionado

  • claude-terminal-mcp — terminal, sistema de archivos y trabajos en segundo plano.

  • claude-linux-mcp — control de escritorio X11 (captura de pantalla, ratón, teclado, ventanas, portapapeles).

Available Tools

5 tools
gpu_metricsA
Read-only

Full rocm-smi -a --json output for every GPU (clocks, voltages, PCIe link width/speed, firmware versions, per-engine activity, throttle status, energy counters). Use when gpu_status is not enough. The shape is rocm-smi’s native JSON, unmodified.

ParametersJSON Schema
NameRequiredDescriptionDefault

No parameters

TDQS

A4.7/5.0
Behavior4/5

Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?

Annotations already mark readOnlyHint=true and destructiveHint=false. The description adds context about the tool being a full dump (clocks, voltages, PCIe, firmware, etc.) and that the output is unmodified, which is helpful beyond annotations.

Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.

Conciseness5/5

Is the description appropriately sized, front-loaded, and free of redundancy?

Two sentences, front-loading the core purpose and details. Every sentence adds value: first defines tool, second provides usage guidance. No wasted words.

Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.

Completeness5/5

Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?

Given no parameters and no output schema, the description sufficiently explains what the tool returns (full rocm-smi -a --json output) and when to use it. Annotations complement the read-only nature. Complete for a simple metrics tool.

Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.

Parameters4/5

Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?

The input schema has no parameters, and schema coverage is 100%. With zero params, the description does not need to add parameter info. Baseline of 4 is appropriate.

Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.

Purpose5/5

Does the description clearly state what the tool does and how it differs from similar tools?

The description clearly states the tool returns full rocm-smi -a --json output for every GPU with specific fields listed. It distinguishes from sibling tools by indicating when to use it (when gpu_status is not enough).

Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.

Usage Guidelines5/5

Does the description explain when to use this tool, when not to, or what alternatives exist?

Explicitly states 'Use when gpu_status is not enough', providing clear guidance on when to prefer this tool over alternatives. Also notes the output shape is unmodified rocm-smi JSON.

Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.

gpu_processesA
Read-only

List compute processes using the GPU (KFD PIDs) with their VRAM usage and card index. Returns an empty list when no compute workloads are running.

ParametersJSON Schema
NameRequiredDescriptionDefault

No parameters

TDQS

A4/5.0
Behavior4/5

Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?

Description adds behavioral context beyond annotations (e.g., returns empty list when no workloads) and confirms read-only nature. No contradictions.

Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.

Conciseness5/5

Is the description appropriately sized, front-loaded, and free of redundancy?

Two succinct sentences covering purpose and behavior, front-loaded with the action, no unnecessary words.

Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.

Completeness4/5

Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?

Without output schema, the description outlines return content (compute processes, VRAM, card index) and edge case. Could specify field details, but adequate for such a simple tool.

Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.

Parameters4/5

Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?

No parameters exist, so baseline 4 applies. Description need not add parameter info.

Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.

Purpose5/5

Does the description clearly state what the tool does and how it differs from similar tools?

The description uses specific verb 'List' and resource 'compute processes using the GPU (KFD PIDs)', clearly distinguishing it from siblings like gpu_metrics or gpu_status.

Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.

Usage Guidelines2/5

Does the description explain when to use this tool, when not to, or what alternatives exist?

No guidance on when to use this tool vs alternatives (e.g., gpu_watch, rocm_info), and no exclusions or prerequisites mentioned.

Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.

gpu_statusA
Read-only

One-shot summary of all AMD GPUs: product name, GPU utilization %, VRAM used/total bytes and %, edge/junction/memory temperatures, average and max power, fan speed % and RPM. Returns one entry per card. Fields that the card does not support are returned as null (not omitted).

ParametersJSON Schema
NameRequiredDescriptionDefault

No parameters

TDQS

A4.1/5.0
Behavior3/5

Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?

Annotations already declare readOnlyHint=true and destructiveHint=false. Description adds that unsupported fields return null, which is useful but not extensive. No contradiction.

Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.

Conciseness5/5

Is the description appropriately sized, front-loaded, and free of redundancy?

Single concise sentence listing all fields. No wasted words.

Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.

Completeness5/5

Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?

Despite no output schema, description explains return structure (one entry per card, null for unsupported fields). Complete for the tool's simplicity.

Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.

Parameters4/5

Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?

No parameters, schema coverage 100% trivially. Description doesn't add parameter meaning, but none needed. Baseline 4 is appropriate.

Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.

Purpose5/5

Does the description clearly state what the tool does and how it differs from similar tools?

Description clearly states it provides a 'one-shot summary of all AMD GPUs' and lists specific fields, distinguishing it from siblings like gpu_metrics which likely provide more detailed historical data.

Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.

Usage Guidelines3/5

Does the description explain when to use this tool, when not to, or what alternatives exist?

No explicit when-to-use or alternatives mentioned. The description implies it is for a quick overview, but does not guide against using other tools for specific needs.

Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.

gpu_watchA
Read-only

Take N snapshots of gpu_status at a fixed interval and return both the raw frames and per-card min/max/avg statistics for utilization, temperature, power, and VRAM usage. Useful for answering “is this training run stable?”. Default: 5 samples at 1000ms intervals.

ParametersJSON Schema
NameRequiredDescriptionDefault
samplesNoNumber of samples to take (2–60). Default: 5.
interval_msNoMilliseconds between samples (100–10000). Default: 1000.

TDQS

A4/5.0
Behavior3/5

Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?

Annotations already declare readOnlyHint=true, so no contradiction. Description adds that it returns raw and aggregated data, but omits that the tool blocks for the total sampling duration, which is important for agent planning.

Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.

Conciseness5/5

Is the description appropriately sized, front-loaded, and free of redundancy?

Two sentences with no fluff: purpose, use case, and defaults are front-loaded efficiently.

Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.

Completeness4/5

Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?

Given 2 params with full schema coverage, no output schema, and sibling tools, the description adequately explains input, output (raw and stats), and use case. Lacks return structure details but acceptable.

Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.

Parameters3/5

Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?

Schema coverage is 100% with clear parameter descriptions including defaults. The tool description does not add meaning beyond what the schema already provides, so baseline 3 is appropriate.

Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.

Purpose5/5

Does the description clearly state what the tool does and how it differs from similar tools?

The description clearly states it takes N snapshots of gpu_status at intervals and returns raw frames plus per-card statistics, distinguishing it from siblings like gpu_status (single snapshot) and gpu_metrics (single metrics).

Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.

Usage Guidelines4/5

Does the description explain when to use this tool, when not to, or what alternatives exist?

It provides a specific use case ('is this training run stable?') and default parameters, implying when to use it. However, it does not explicitly mention alternatives or when not to use.

Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.

rocm_infoA
Read-only

Report the rocm-smi version, kernel driver version, whether the amdgpu module is loaded, installed ROCm/HIP/HSA packages (from dpkg), and whether amdgpu_top is available. Useful for checking ROCm install health before running workloads.

ParametersJSON Schema
NameRequiredDescriptionDefault

No parameters

TDQS

A4.4/5.0
Behavior4/5

Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?

Annotations already declare readOnlyHint=true and destructiveHint=false. The description adds detail on exactly what is reported (versions, packages, etc.), providing behavioral context beyond the annotations.

Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.

Conciseness5/5

Is the description appropriately sized, front-loaded, and free of redundancy?

Single sentence efficiently lists all reported information and includes a usage note. Every word earns its place.

Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.

Completeness4/5

Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?

For a simple info tool with no output schema, the description adequately covers purpose and content. Could optionally mention return format, but not necessary.

Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.

Parameters4/5

Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?

Tool has zero parameters, so schema covers all. Description adds no parameter info, but this is acceptable given no params. Baseline 4 applies.

Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.

Purpose5/5

Does the description clearly state what the tool does and how it differs from similar tools?

The description uses the specific verb 'Report' and lists concrete items (rocm-smi version, kernel driver, etc.), clearly distinguishing this from sibling tools that focus on metrics, processes, or status.

Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.

Usage Guidelines4/5

Does the description explain when to use this tool, when not to, or what alternatives exist?

States explicit use case: 'checking ROCm install health before running workloads.' This guides the agent on when to invoke. While it doesn't explicitly mention alternatives, the context is clear enough.

Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

  1. 5 tool updatesv0.1.0
    • First observedgpu_metrics
    • First observedgpu_processes
    • First observedgpu_status
    • First observedgpu_watch
    • First observedrocm_info
Deploy Server
Shareable, User-Owned Context BackendPeople, applications, and agents read, write, and coordinate together in one context lake. Reach it over MCP, a CLI, a Python SDK, or REST.fulcradynamics.comAd

TDQS

A4.3/5.0

Scored across 5 tools

Disambiguation5/5

Each tool serves a distinct purpose: raw metrics, summary status, process listing, time-series snapshots, and install health check. No overlap in functionality.

Naming Consistency5/5

All tool names follow a consistent snake_case pattern with a clear noun-based prefix (gpu_ or rocm_), making them predictable and easy to distinguish.

Tool Count5/5

Five tools is appropriate for a monitoring-focused server, covering essential needs without being excessive or insufficient.

Completeness4/5

The tool set covers core monitoring, process insight, and installation health, but lacks a way to reset or configure GPU parameters, though that may be out of scope.

Maintenance

ActivityMaintained
ResponsivenessNo issues

Resources

Related MCP Connectors

Related MCP Servers

  • System Monitor MCP Server

    PraveerT
    F
    license
    Not graded
    quality
    D
    maintenance
    Provides system monitoring and management capabilities for Claude CLI, allowing users to view system information, track resource usage, and manage processes through natural language commands.
    -

  • mcp-system-monitor

    VZilvinskus
    F
    license
    A
    quality
    D
    maintenance
    Enables Claude to monitor CPU, memory, disk, processes, network connections, Docker containers, and system logs for system diagnostics and troubleshooting.
    5
    -