rajudandigam/agent-inspect

What it solves

Agent code is complex and rarely fails in a single function call; it involves planning, tool use, and retries. Flat logs make it difficult to visualize the execution path and identify the first causal failure. AgentInspect provides a local-first approach to capture agent runs as readable execution trees, allowing developers to debug, verify trajectories in CI, and share redacted evidence without needing a hosted service.

How it works

It captures agent runs as local JSONL files. It provides a CLI to view these traces as execution trees, including nested steps, tool calls, and LLM metadata. It allows developers to define "trace contracts" (deterministic rules) to check if an agent followed the correct trajectory (e.g., ensuring a specific tool was called). These traces can be redacted and bundled into portable, hash-verifiable evidence artifacts for sharing.

Who it’s for

TypeScript developers building AI agents using frameworks like LangChain, LangGraph, AI SDK, or OpenAI Agents JS, as well as those using custom agentic workflows.

Highlights

• Local-first: No account, collector, or default upload required; traces stay on the developer's machine.
• Deterministic Trajectory Checks: Fail CI builds based on specific tool calls or completion statuses.
• Evidence Bundling: Create redacted, offline artifacts for PRs or incident reports.
• Broad Integration: Supports multiple TypeScript AI frameworks and standard formats like OpenInference/OTLP JSON.
• MCP Server: Optional support for exposing local evidence to MCP clients like Cursor.