08/06/2026
๐๐ ๐๐ฎ๐ข๐ฅ๐ญ ๐ ๐ฌ๐๐ฅ๐-๐ก๐จ๐ฌ๐ญ๐๐ ๐๐๐๐ฉ ๐ซ๐๐ฌ๐๐๐ซ๐๐ก ๐๐ ๐๐ง๐ญ ๐จ๐ง ๐๐ฅ๐๐ฎ๐๐ ๐๐จ๐๐. ๐๐ญ'๐ฌ ๐๐๐ฅ๐ฅ๐๐ ๐๐๐จ๐ฎ๐ญ.
The problem: research ideas hit on the train, in line, mid-conversation. You'd start a thread in Claude web, then lose it โ or find a Q&A scroll with no synthesis. Useless as context for actual work.
What we wanted: kick off research from a phone (a GitHub Issue), come back to the desktop with the synthesis ready. Bonus: by then your Pro quota has reset.
๐๐จ๐ฐ ๐ข๐ญ ๐ฐ๐จ๐ซ๐ค๐ฌ:
โ Open a GitHub Issue with your topic
โ Scout sharpens it (expands your initial query)
โ Wide topics decompose into parallel sub-agent expeditions
โ Each expedition runs independently, then a synthesis pass reconciles everything
โ Published as markdown (and/or fancy html view) on a Jekyll site you own
โ Uses your existing Pro/Max quota โ no extra API bill
๐
๐จ๐ฎ๐ซ ๐๐๐๐ข๐ฌ๐ข๐จ๐ง๐ฌ ๐ญ๐ก๐๐ญ ๐ฌ๐ก๐๐ฉ๐๐ ๐ข๐ญ:
๐๐ข๐ญ๐๐ฎ๐ ๐๐ฌ๐ฌ๐ฎ๐๐ฌ ๐๐ฌ ๐ญ๐ก๐ ๐๐ โ Auth, comment threads, checkboxes. We didn't have to build anything.
๐๐ก๐๐ซ๐ฉ๐๐ง ๐๐๐๐จ๐ซ๐ ๐ซ๐๐ฌ๐๐๐ซ๐๐ก โ "Best NAS to buy in 2026" is three questions in a trench coat. The sharpening step proposes a tight framing before any research starts.
๐๐๐๐จ๐ฆ๐ฉ๐จ๐ฌ๐ ๐ข๐ง๐ญ๐จ ๐ฉ๐๐ซ๐๐ฅ๐ฅ๐๐ฅ ๐๐ฑ๐ฉ๐๐๐ข๐ญ๐ข๐จ๐ง๐ฌ โ Map-reduce at the research-question level. Sub-topics must be independently coherent. The synthesis pass is where most value lives.
๐๐ฏ๐๐ซ๐ฒ ๐๐ฅ๐๐ข๐ฆ ๐๐๐ซ๐ซ๐ข๐๐ฌ ๐ข๐ญ๐ฌ ๐๐๐ ๐ข๐ง๐ฅ๐ข๐ง๐ โ No "References" dump at the end. `[[1]](url)` next to every claim makes it locally auditable: one click verifies one sentence.
MIT-licensed. One-line install. Three repos (Scout / Atlas / Compass) โ fork it, take the patterns, make it your own.
Repo: https://github.com/Laoujin/Scout
Live example: https://laoujin.github.io/Atlas/
Full blog post: https://itenium.be/blog/ai/scout-a-self-hosted-deep-research-agent-on-claude-code-2/