AgentInspect

AgentInspect records TypeScript and Node.js agent runs as local traces and uses each trace for three jobs: debugging the run, failing CI when the agent takes the wrong path, and sharing a redacted record of what happened. It needs no account, no collector, and uploads nothing by default.

A run is stored as local JSONL. The viewer turns it into an execution tree of nested steps, tool calls, LLM calls, model and token metadata, durations, errors, and the first causal failure, and can produce a summary report or an explanation of the run. Traces come from manual instrumentation (wrapping steps, tool calls, and LLM calls), from adapters for LangChain and LangGraph, the AI SDK, and OpenAI Agents JS, from existing structured logs, or from OpenInference and OTLP JSON files.

Trajectory checks are deterministic and use no LLM judge: the same trace and rules always give the same result. You start from a preset and add expectations such as a required tool, then run the check as a CI gate that exits non-zero on failure. TraceContract defines the same rules in code, and Vitest and Jest reporters, suites, and cohorts cover larger test setups.

For sharing, a safety check rates a run as safe, safe with warnings, unsafe, or unknown, and a bundle command writes a redacted Evidence v2 artifact with a manifest whose file hashes can be verified offline. Capture is metadata-only by default, so raw prompts and outputs are recorded only on opt-in, and redaction runs at write time. An optional, read-only MCP server exposes local evidence to coding agents. AgentInspect is not a hosted observability service, a production trace store, or an LLM-judge evaluation platform, and it is meant to sit alongside those.