AgentSight is powerful because it observes sensitive boundaries. Treat the traces accordingly.
System-level profiling can capture prompts, responses, paths, headers, commands, network destinations, and resource data. The product is designed for local-first workflows and explicit exports, not hidden hosted collection.
Local-first data path
Recorded sessions are saved locally unless you explicitly export or serve them. Captured data can include prompts, responses, paths, headers, and network targets.
Targeted tracing
AgentSight is designed around recording the selected agent command, process, or session rather than turning the product site into a hosted collector.
No SDK, no proxy, no vendor hooks
The profiler observes stable system boundaries, so it can work across runtimes and closed-source CLIs where application traces stop.
OTel-compatible output
Captured LLM calls can be exported as OpenTelemetry GenAI spans for teams that already operate standard telemetry pipelines.
Operational notes
AgentSight needs Linux eBPF support and elevated privileges for live capture. This is the tradeoff that makes external process, file, network, and TLS boundary profiling possible.
Saved sessions should be handled like logs containing source code, prompts, headers, local paths, and infrastructure metadata. Use local storage, redaction, or controlled export workflows before sharing reports.
The live demo at app.agentsight.us uses a recorded sample session. It is separate from this product site so the marketing page remains static, crawlable, and easy to deploy.