Labsco
MCP SERVER

MCP Agent Trace Inspector

by dbsectrainer

Record what an agent actually did — every tool call, its latency and its token count — into a local SQLite file you can read back.

LLM Evaluation & Observability
Summary
Observability that stays on the machine.

The trade is deliberate: no account, no API key and no traces leaving the laptop, in exchange for no cross-team sharing. Where the tool outputs are sensitive, that is the right way round.

What it is

A local-first tracing layer for agent workflows. A trace is opened, steps are recorded against it as the work proceeds, and it is closed; everything lands in a SQLite database on your machine and nothing is sent anywhere.

What you get
  • Every step recorded with its inputs, outputs, latency and token usage
  • Traces that survive a restart, stored in local SQLite with no external dependency
  • A self-contained HTML dashboard with an interactive step timeline
  • Cost per trace estimated offline from a pricing table you configure, with no API traffic required
  • Two traces diffed side by side, to measure what a prompt or tool change actually did
  • Under 5ms added per step
  • Past traces listed and reopened
Requirements

Node with npx — the package runs without installing and writes its database locally. Traces hold the inputs and outputs your tools saw, so the database file is as sensitive as the data that passed through it.

Setup effort

One command — npx -y mcp-agent-trace-inspector@latest