Ingero

If you're debugging GPU workloads and want Claude to tell you why your training job stalled instead of staring at metrics dashboards, this is the bridge. It exposes seven MCP tools that wrap Ingero's eBPF tracing engine, letting your AI assistant query causal chains that connect Linux scheduler events, CUDA Runtime and Driver API calls, and Python source lines. Ask "what caused the GPU stall" and get back a timeline showing cudaMalloc spikes during CPU contention, complete with the exact forward() call and line number. Works with PyTorch and TensorFlow out of the box, traces both Runtime and Driver layers to catch cuBLAS and torch.compile kernels most profilers miss, and runs with under 2% overhead. No code changes, just point it at a running process or query historical traces through the MCP interface.