Open Source · Live Demo

An agent trace you can actually see

An infinite, Miro-style canvas that renders an AI agent's run as a node-edge graph — every model call, tool call and retrieval, each scored by an eval. The demo below replays a real run of the PubCrawl literature agent answering whether lecanemab slows cognitive decline in early Alzheimer's — watch it draft an answer that overstates the clinical effect, and the eval catch it. Or ask your own biomedical question and watch a real agent search the evidence and answer, live.

Real PubCrawl agent run Per-step eval scoring Live WebSocket replay Run a real agent yourself

View on GitHub Open full screen

TypeScript · built on React, Konva & WebSockets · real data via PubCrawl over PubMed & ClinicalTrials.gov

01Live demo

Watch the agent think

Watch a saved run “What is the evidence for lecanemab slowing cognitive decline in early Alzheimer's?” Watch a finished run play back, step by step.

⚡ Ask your own Watch it search the literature and answer in real time.

Use the controls above to replay the saved run or ask your own question. On the canvas: drag to pan, scroll to zoom, click any step to inspect its input, output and eval. Replay uses real tool inputs/outputs (PubCrawl over PubMed & ClinicalTrials.gov) with illustrative timings; Run live executes a real Claude Sonnet 4.6 + PubCrawl agent.