# AgentInspect

A local-first debugging and CI toolkit for TypeScript AI agents that turns raw execution traces into readable execution trees, deterministic trajectory checks for catching wrong tool paths, and shareable offline evidence bundles.

AgentInspect captures a TypeScript agent's run as local JSONL and reuses that one trace for three purposes. For debugging, it renders nested steps, tool calls, model calls, durations, errors, and the first causal failure without sending anything to an external service. For CI, its trajectory checks are deterministic and provider-free -- the same trace and rules always produce the same result -- and can require a specific tool to have been called or fail when an explicit outcome event is recorded; a passing check exits with success and a rule failure exits with an error code. For sharing, dedicated commands produce a reviewable, offline evidence package that can later be checked for integrity.

It works with manual instrumentation, official adapters, or logs already emitted by frameworks like LangChain, LangGraph, or Vercel AI, and needs no account, collector, or default upload.

It installs with npm and is run through its own CLI.

Website: https://agentinspect.vercel.app