Reactor - Self-hosted agent that researches across walled gardens

by
一个面向全网研究与复杂数据分析的云端智能体,能在公开网络、YouTube、BiliBili、X、小红书、Hacker News上研究任何主题 。项目wiki: - OWWZO/ai-agent

Add a comment

Replies

Best
Maker
📌
Hey Product Hunt — I'm the maker of Reactor. I built this because a research question still lives in six gardens. Reddit threads, X, YouTube transcripts, HN, plus Bilibili / Xiaohongshu / V2EX for CN topics. Search APIs mostly don't see them. Chat products each cover one fence. Reactor is open source and self-hosted. You give a goal; it plans, calls tools, checks results, and revises on failure. Vague ask → it stops and asks. Plan Mode: until you approve, it may only research and write the plan — no DB writes, no code, no report. Hard tasks split to sub-agents (wait, or background). Tools: web search, RAG, sandbox code, charts, docs. NL2SQL uses Table RAG, then previews SQL. Collectors run in parallel so community threads and public pages share one citation list. Private PDF/Word/PPT/images are parsed/OCR'd and cited the same way. MCP tools stay as names until the agent asks for a schema. Memory is not the whole chat dumped into the prompt. Facts inject every turn; history is retrieved on demand; reusable workflows live as SKILL.md. Output lands in the thread (charts, tables, canvas) or exports to PDF/Word/PPTX/HTML. What I actually use it for: 1. Brief before a meeting — last ~30 days of public posts, cited 2. Data analysis / a simple backtest 3. Competitor teardown with high-vote comments 4. Trip notes from Xiaohongshu/Bilibili Stack: Java 21 + Spring AI harness, Python sidecar, React workbench. Docker Compose in the repo. Bring your own LLM endpoint (local or remote). Repo: Would love honest takes on Plan Mode and the Java/Python split — not looking for upvotes, looking for "this part is wrong."