Trace agent runs. Guard proposed tool calls.
Use the Python SDK for non-streaming OpenAI Chat Completions, Anthropic Messages, or direct LiteLLM calls on the tested OpenAI route. Connect Claude Code or Codex with one setup command.
Start with the launch quickstart
Launch quickstart
Run the pinned OpenAI sample and check the exact trace shape.
OpenAI Chat Completions
Instrument a non-streaming Chat Completions loop.
Anthropic Messages
Instrument a non-streaming Messages loop.
LiteLLM Python
Instrument direct non-streaming completion calls on the tested OpenAI route.
Python decorators
Define the agent root and record real tool execution.
Claude Code
Connect Claude Code hooks and review captured sessions.
Codex
Connect Codex hooks and review captured sessions.
What the launch path covers
Observe
Capture explicitly decorated agent, LLM, and tool spans with real parent-child links.
Guard
Start configured policies in audit mode, then enforce blocking when ready.
Provider-reported token usage is recorded when present. Cost is model-dependent and may be unavailable. capture_messages=False suppresses the request and system_prompt categories. It does not redact other captured fields.
For setup details, see SDK setup. For policy behavior, see Guard rules and policies.
For application-specific conditions and authenticated decision services, see Python rules.