Getting Started¶

Prerequisites¶

Go 1.23+
CGO enabled
NVIDIA drivers and NVML headers (Linux, for real GPU access)

Tests use a mock collector and run anywhere — no GPU hardware required.

Build¶

make build       # compile binary (requires CGO + NVML on Linux)
make test        # run tests (no GPU needed — uses mock)
make lint        # golangci-lint
make docker      # container image

Run¶

# stdio transport (default) — the agent spawns this as a child process
./gpu-mcp-server

# check version
./gpu-mcp-server --version

The server communicates over stdio using the MCP protocol. It reads GPU state from NVML on demand — no polling, no daemon.

Docker¶

make docker
docker run --gpus all --rm gpu-mcp-server:dev

Helm (Kubernetes)¶

helm install gpu-mcp-server deploy/helm/gpu-mcp-server

Verify¶

# run the test suite (works with no GPU)
make test

# check code quality
make lint
make vet
make fmt