Launching today
WorkloadTruth
Verify if a GPU job is training or idling
1 follower
Verify if a GPU job is training or idling
1 follower
WorkloadTruth is an Apache 2.0 CLI and MCP server that classifies GPU workloads as training, inference, or idle entirely from telemetry like utilization, memory, and power. No code changes or self-reporting are needed. We ship a hash-chained audit log and an evasion benchmark. Unlike NVIDIA DCGM that exposes raw metrics or run:ai that trusts user labels, our approach classifies actual behavior.
As AI infrastructure scales, gaining true visibility into GPU clusters has become a massive headache. Most monitoring dashboards give you raw, out-of-context metrics. You might see a GPU pinned at 100% utilization, but you have no idea what it is actually doing. We were tired of playing a guessing game to figure out if a node was crunching through a training job or serving user inference requests.
We built WorkloadTruth to eliminate this blind spot. It is designed to classify GPU workloads strictly as either inference or training based solely on telemetry data. By looking at hardware behavior signatures (rather than application code), you get the ground truth of what your GPUs are doing.
Initially, we were going to build this as a standalone monitoring script or a traditional DevOps dashboard widget. But while building the classification logic, we realized that the modern developer workflow is shifting towards AI agents. Instead of making users check yet another dashboard, we evolved the project into a Model Context Protocol (MCP) server.
By exposing these classification capabilities via MCP tools, you can now seamlessly connect WorkloadTruth to clients like Claude Desktop or Cursor. Now, you can just ask your AI assistant what your GPUs are doing, and it can natively read the telemetry to give you an answer.
Repo: https://github.com/RudrenduPaul/WorkloadTruth
MCP Servers:
https://mcpservers.org/servers/rudrendupaul/workloadtruth
https://glama.ai/mcp/servers/RudrenduPaul/workloadtruth
NPM: https://www.npmjs.com/package/workloadtruth-cli
PyPI: https://pypi.org/project/workloadtruth-cli
We’d love to hear your feedback, feature requests, or how you plan to integrate this into your own AI infrastructure.