An autonomous triage agent for open-source repos. Give it owner/repository,
it reads the open issues & PRs, classifies each one with an LLM
(bug / feature / question / duplicate / docs), assigns a priority (1–5),
and renders an at-a-glance dashboard with suggested next actions.
Built as a submission demo for online agent hackathons (All Things Agentic / Agents for Humans), and as a practical tool for maintainers.
- Fetch — GitHub REST API lists open issues + PRs (public).
- Reason — each item is sent to an LLM (OpenRouter free tier, model
rotation with fallback) with a strict JSON output contract:
{"category", "short", "priority"}. - Render — a static-first page (no build step) shows cards sorted by priority, color-coded by category, each linking to the original issue.
Everything runs on demand — no background jobs, no servers except a thin local proxy for the LLM call & GitHub rate limits.
export LLM_KEY="<openrouter key>"
node server.mjs # serves :7811
open http://localhost:7811server.mjs— minimal Node HTTP server + agent loop + LLM rotationindex.html— single-file UI (no dependencies, inline CSS/JS)
Node.js (bare node:http), GitHub REST API, OpenRouter :free models,
no frameworks, no build step — total ~250 lines.
typeorm/typeorm— 22 items scanned, classified (bug/feature/docs…)facebook/react— 22 items scanned, 22/22 classified, priorities p1–p5 (e.g. p1 bug: "Freezing props prevents later assignments…", p1 bug: "Unbounded debug info causes RangeError on large…")
MIT — see LICENSE.
