NVIDIA DCGM MCP server: GPU metrics via community servers

DCGM telemetry usually lands in Prometheus through dcgm-exporter. An agent can query it there, or use a community server that inspects GPUs on the node itself.

Checked 3 October 2026 · from NVIDIA's DCGM and dcgm-exporter repositories and docs, the community server repositories, npm and the GitHub container registry

One of 91 guides to monitoring and developer tools. Each covers what an AI coding agent can get out of the tool through MCP, the protocol that lets the agent query the tool directly.

How NVIDIA DCGM works with MCP

DCGM runs on each GPU node. It reads temperatures, clocks, power, memory use, NVLink traffic and XID errors. XID errors are the codes the NVIDIA driver logs when a GPU faults.

dcgm-exporter is NVIDIA's companion that publishes those readings as Prometheus metrics, such as DCGM_FI_DEV_GPU_UTIL for utilization. It has 1,890 GitHub stars and was last pushed on 18 September 2026.

DCGM also runs active diagnostics, which stress a GPU to confirm it is healthy. Google Cloud's Cluster Director MCP server exposes those diagnostics as a single tool.

Every vendor here gets the same four checks: official, maintained, distributed and usable. NVIDIA DCGM passes 1 of them.

DCGM answers whether a GPU is busy, hot or failing. For which job caused the load, pair it with a Kubernetes or scheduler view. For what the GPU hours cost, use a cost tool.

What can an AI agent do with NVIDIA DCGM?

How to configure NVIDIA DCGM MCP for agents

1. 1

   Read-only first

   Prometheus's core tool profile leaves out its administrative tools, which can delete series or shut the server down.

2. 2

   Node access is a wider grant

   k8s-gpu-mcp-server runs on GPU nodes and gpu-watch uses SSH, so both see more than a Prometheus query does.

3. 3

   Keep the operator tools off

   k8s-gpu-mcp-server plans kill_gpu_process and reset_gpu behind an explicit operator mode; its default is --mode=read-only.

Paste this into your agent’s instructions

When you answer from DCGM data, state:
- The metric names, the GPUs or nodes, and the time range.
- Whether the data came from Prometheus or from a node-level tool.
Report an XID value as an error code, not a count.
If no DCGM series exist, say the exporter may not be scraped.

Do you need the MCP server at all?

Prometheus. The prometheus/prometheus-mcp server is first-party, with 115 stars and a push on 2 October 2026. The agent writes PromQL, the Prometheus query language, against the DCGM_ series.

Grafana. If dashboards and alerts already sit on that Prometheus, Grafana's server adds alert rules and panels over the same GPU data.

`ArangoGutierrez/k8s-gpu-mcp-server`. It inspects GPUs on each Kubernetes node through NVML, the NVIDIA Management Library, and uses DCGM when available. Tools include get_gpu_health, analyze_xid_errors and get_pod_gpu_allocation.

`GoogleCloudPlatform/cluster-director-mcp`. For clusters built with Google Cloud Cluster Director, its run_dcgm_test tool runs DCGM diagnostics on the GPU nodes. It is a Gemini CLI extension meant for Cloud Shell.

`01p5/gpu-watch`. A read-only Slurm-oriented monitor that reaches nodes over SSH and parses dcgm-exporter's /metrics endpoint directly. It has 0 stars and was last pushed on 1 June 2026.

Query dcgm-exporter through Prometheus for trends and alerts. Add a node-level community server only when the agent must diagnose a specific failing GPU.

NVIDIA DCGM integration setup

The working route needs no DCGM-specific server. Point Prometheus's MCP server at the Prometheus that scrapes dcgm-exporter.

claude mcp add-json "prometheus" '{"command":
  "prometheus-mcp-server",
  "args":["--mcp.tools=core",
          "--prometheus.url=http://prometheus:9090"]}'

curl -s "http://prometheus:9090/api/v1/label/__name__/values" \
  | grep -o '"DCGM_[^"]*"' | head

avg_over_time(DCGM_FI_DEV_GPU_UTIL[1h])

Every other client

Each block below is the configuration for one client, with the file path and the key that client expects.

What engineers report

Adoption of DCGM MCP servers is small. k8s-gpu-mcp-server has 4 stars, Cluster Director's server has 9, and gpu-watch has none. dcgm-exporter itself has 1,890 stars, which reflects how most teams already consume DCGM data.

Should you connect an agent to NVIDIA DCGM?

Yes, through Prometheus, if your cluster already scrapes dcgm-exporter. Best for platform teams checking GPU utilization, temperature and XID errors across a cluster.

Strongest fit

Yes, through Prometheus, if your cluster already scrapes dcgm-exporter.

Main advantage

GPU telemetry is usually already in Prometheus, and Prometheus has a maintained first-party server.

Main weakness

NVIDIA ships no GPU-aware tools, so the agent works from raw metric series.

Operational risk

Community servers that run on GPU nodes or over SSH hold far wider access than a read-only Prometheus query.

Frequently asked questions