claude-rocm-mcp
Read-only AMD GPU monitoring for Claude on Linux — VRAM, temperature, clocks, power, and running ML processes via rocm-smi.
README
Claude ROCm GPU Monitor
Read-only AMD GPU monitoring for Claude Desktop on Linux. Lets Claude answer questions like "is my training run healthy?", "how hot is the GPU?", "how much VRAM is left?", and "which process is using the card?" — all without the ability to kill processes, change clocks, or otherwise perturb a running workload.
It shells out to rocm-smi (and optionally amdgpu_top) and exposes five strictly read-only tools. No process killing, no clock/power overrides, no fan control. For those, use a tool like gpu-kill.
Requirements
- Linux with an AMD GPU (RDNA2/3/4 consumer cards — RX 6000/7000/9000, R9700 — and CDNA data-center cards).
amdgpukernel module loaded (standard on modern Ubuntu/Debian/Fedora).rocm-smiinstalled:sudo apt install rocm-smi # minimal # or a full ROCm stack: # https://rocm.docs.amd.com/projects/install-on-linux/
Some metrics (clocks, power, detailed temperatures) require newer ROCm releases. If a metric isn't supported by your card/driver, the tool returns it as null rather than failing.
Install (Claude Desktop)
- Download the latest
ROCm.mcpbfrom the Releases page. - In Claude Desktop, open Settings → Extensions.
- Scroll to Extension Developer at the bottom, click Install Extension, and pick the
.mcpbfile. - Enable the extension. The tools appear prefixed with
rocm-gpu-monitor:in Claude's tool picker.
Tools
| Tool | What it does |
|---|---|
gpu_status |
One-shot summary: name, utilization %, VRAM used/total, temps (edge/junction/memory), power avg/max, fan %/RPM. |
gpu_metrics |
Full rocm-smi -a --json dump — clocks, voltages, PCIe link, firmware versions, per-engine activity. |
gpu_processes |
Compute processes using the GPU (KFD PIDs) with VRAM usage and card index. |
gpu_watch |
Take N samples at a fixed interval and return raw frames plus per-card min/max/avg stats. |
rocm_info |
Installed ROCm/HIP packages, driver version, amdgpu module load status, amdgpu_top availability. |
All tools carry the MCP readOnlyHint: true / destructiveHint: false / openWorldHint: false annotations.
Example prompts
"Is my training run stable? Watch the GPU for 20 seconds."
"How much VRAM is my current PyTorch process using?"
"Is the GPU thermal-throttling? What's the junction temperature?"
"What ROCm version is installed and is the driver loaded?"
Privacy policy
This extension runs entirely on your local machine and shells out only to the following programs, always read-only:
rocm-smi— queries the AMD kernel driver for device metrics.amdgpu_top(if installed) — optional richer metrics.dpkg -l— enumerates installed ROCm-related packages by name/version.lsmod— checks whether theamdgpukernel module is loaded.which— locates the above binaries at startup.
No data leaves your machine. This extension performs no network I/O, opens no sockets, and writes no files outside of standard process stdout/stderr (captured by Claude Desktop's log directory). It cannot modify GPU state (clocks, power, fan), kill processes, or otherwise perturb a running workload — every tool is strictly read-only.
The information visible to Claude includes GPU model and firmware versions, current VRAM usage, temperature/clock/power readings, running GPU-using process PIDs and their VRAM usage, and a list of installed ROCm-related packages. If you consider any of that sensitive (for example the list of running process PIDs on a shared machine), do not enable this extension.
Troubleshooting
"rocm-smi is not installed" — install it (sudo apt install rocm-smi). If you're using a non-Debian distro, follow AMD's ROCm install instructions.
Many fields come back as null on a new GPU — some metrics (clocks, power, temperature) depend on firmware support exposed through rocm-smi. On very new cards (e.g. RDNA4/R9700), individual rocm-smi --showclocks/--showpower may report "No JSON data to report" even though -a --json returns the data. This extension uses -a --json as the primary source to maximize coverage. If a specific field is unsupported on your card, it'll be null rather than an error.
Permission errors — rocm-smi normally runs without root. If you see permission errors, check that the user running Claude Desktop is in the render and video groups:
sudo usermod -aG render,video $USER
# log out and back in
Development
The server is a single ~300-line Node.js script with zero npm dependencies. To rebuild the .mcpb:
cd bundle-source
zip -j ../ROCm.mcpb manifest.json package.json server.js README.md LICENSE icon.png
License
MIT. See LICENSE.
Related
- claude-terminal-mcp — shell, filesystem, and background jobs.
- claude-linux-mcp — X11 desktop control (screenshot, mouse, keyboard, windows, clipboard).
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。