Vibe Coding Discoverflameox
View on GitHubRuntime evidence that helps agents trace, profile, and burn down hotspots in application and native code, GPU kernels, and inference stacks.
Flameox is a local-first MCP server that lets coding agents capture and analyze bounded runtime evidence: CPU/GPU profiles, traces, memory data, and benchmark comparisons across Python, native code, and inference stacks. It coordinates py-spy, Memray, Perfetto, OTLP, torch, and aiperf with explicit artifact paths and n
Use Cases
Profile Python applications from an MCP-capable coding agentBurn down GPU kernel hotspots from existing artifactsCompare baseline vs candidate benchmarks for performance regressionsAnalyze LLM inference stack performance with aiperfTrace and inspect OTLP/Perfetto tracesInvestigate memory usage with MemraySample CPU stacks with py-spyCapture bounded runtime evidence for a live commandPreserve immutable, SHA-256-bound evidence bundlesQuery previously captured evidence via MCP resourcesGive agents runtime evidence instead of guessing at bottlenecks
Built With
- Language
- Python
- Frameworks
- Model Context Protocol SDK · Typer · Pydantic · DuckDB · PyArrow · uv/uvx · pytest · Ruff · mypy · py-spy · Memray · Perfetto · OpenTelemetry · PyTorch · coverage.py · aiperf
Tags
mcp · mcp-server · profiler · profiling · gpu-profiling · performance-analysis · benchmarking · runtime-evidence · coding-agents · local-first · tracing · memory-profiling · performance-regression · debugging · python · developer-tools