Debug AI agents locally without a hosted collector

What it does

AgentInspect instruments TypeScript agent runs and writes them as local JSONL traces under .agent-inspect/. It renders nested execution trees for debugging, runs deterministic trajectory checks against those traces in CI, and bundles redacted, hash-verifiable evidence artifacts you can attach to a PR or incident report. The whole loop stays on your machine unless you explicitly decide to share something.

The interesting bit

Most observability tools treat your laptop as a temporary bus stop on the way to their cloud; AgentInspect treats the local file as the final destination. It is deliberately not a SaaS, not a hosted dashboard, and not an LLM-as-judge platform—just a CLI and a set of provider-free checks that fail builds when an agent takes the wrong path or skips a required tool.

Key highlights

• Runs fully offline: core capture, inspection, and checks need no account, no collector, and no default upload.
• Deterministic CI checks: trajectory presets and custom TraceContract rules exit with documented codes, so CI can fail on structural agent mistakes without calling an LLM judge.
• Metadata-only by default: raw prompts and outputs require opt-in, and write-time redaction is enabled out of the box.
• Broad framework support: official adapters for LangChain/LangGraph, Vercel AI SDK, and OpenAI Agents JS, plus ingestion for existing structured logs and OpenInference/OTLP JSON.
• Optional MCP server (Preview) exposes local evidence to coding agents through read-only tools, though the core workflow ignores it.

Caveats

• verify-safe and redaction are best-effort controls, not privacy, security, or compliance certification.
• The VS Code extension exists in the repo but is not yet published to the Marketplace.
• Optional surfaces like the MCP server and Studio are still Preview or Experimental; the stable surface is the core CLI and trace readers.

Verdict

Worth a look if you are building TypeScript agents and want to debug trajectories or gate CI without adopting another hosted observability bill. Skip it if you need fleet-wide production dashboards, hosted trace retention, or a managed prompt registry—it explicitly refuses to be those things.