ppazosp

Agrex

2 min read

Agrex is a real-time graph visualizer for AI agent execution flows. You instrument your agent, and Agrex draws the run as a live graph: agents, tools, and files as nodes, edges showing who spawned what and what got read or written.

agrex.ppazosp.dev · GitHub

Why

Agent systems get hard to reason about fast. Once a run involves tools, files, subagents, and streaming output, logs still tell you what happened, but they hide the shape of the run: which agent called which tool, what fanned out in parallel, where things got stuck. Agrex makes that shape visible.

How it works

Everything is built around a small trace format. A tracer records events as the agent runs. Each event carries fields like parentId, reads, and writes, and Agrex generates the graph edges from those automatically, so you never wire relationships by hand. Nodes update live as work moves from running to done.

There are three ways to use it:

  • @ppazosp/agrex, a React component built on React Flow. Embed it in your own UI and feed it events in real time.
  • agrex on PyPI, a Python tracer that mirrors the JS API. Same trace format, same viewer.
  • The web viewer: drop a JSON or JSONL trace file and scrub through the run. No backend, no setup.

Quick start

# JavaScript / React
npm install @ppazosp/agrex @xyflow/react
 
# Python
uv add agrex

Create a tracer, register your agents, wrap their work in spans, and export with .toJSONL(). The repo has integration examples for the Vercel AI SDK, the Anthropic SDK, the OpenAI SDK, and LangChain.

Layout engines and theming are pluggable. MIT licensed.

Graph View

Backlinks