A research library
you actually own.

Open a GitHub issue. Claude researches, cites every claim, and publishes to a Jekyll site at a URL you own.

Your Claude subscription Your Docker Your URL Open source

Requires: Claude Code, Docker, GitHub account, always-on machine.

Install → See an example Atlas
Scout — the fox mascot standing on a map with compasses overhead
Two repos, one flow

Scout researches. Atlas publishes.

Scout
The researcher

Scout

GitHub Action in your Docker container with Claude Code sharpens your research topic, then starts a recon, survey or expedition.

git push
Atlas
The shelf

Atlas

Publishes Scout's findings to GitHub Pages with your layout configuration.

What sets it apart

Built for serious research.

Identity profile

Localized sharpening

Optional profile.yml with your location, currency, languages, and interests used during sharpening

Depth tiers

Recon / Survey / Expedition

Pick scope per Issue. recon for a quick one-pager, expedition for a thorough multi-angle deep dive.

Decomposition

Wide topics split themselves

Multi-angled requests are decomposed into sub-investigations the user can edit before launch.

Alternative views

Beyond markdown

Markdown by default. Switch to HTML when the output deserves charts, side-by-sides, or layouts text can't reach.

How it works

From issue to Atlas.

YOU OPEN A GITHUB ISSUE SCOUT · YOUR HARDWARE 01 · SHARPEN Bot proposes a sharpened topic tick ✓ · reply to revise · uses your profile · subtopics if wide revise 02 · RESEARCH Claude Code researches & cites every claim WebSearch · WebFetch · Playwright · 5–30 min git push ATLAS · GITHUB PAGES 03 · BUILD GitHub Pages rebuilds Jekyll · markdown or HTML views 04 · PUBLISH You read at your URL laoujin.github.io/Atlas/
Configure

Your Atlas.

Pick a palette, layout, and card — your config s1.rust.v1 feeds the install command below.

$ curl -fsSL https://raw.githubusercontent.com/Laoujin/Scout/main/install.sh \
| bash -s -- --config=s1.rust.v1

Don't pipe strangers into bash? Follow INSTALL.md instead →

Palette · Warm rust

Layout · Hero + footer

Card style · Image-top
https://your-username.github.io/Atlas/
FAQ

Questions you may have.

What does a run cost?

On a Pro/Max subscription you'd already have for coding, runs consume your existing quota — no incremental cost. On API-key auth, measured per-run from my own Atlas (Sonnet 4.6):

TierModeTimeCost median (range)n
reconone-page~2 min$0.83 ($0.73–$0.93)4
surveysingle page5–10 min$2.56 ($1.99–$3.99)24
expeditionsingle deep12–20 min$8.64 ($7.23–$14.90)16
expeditionmulti-angle35–70 min$28 ($12–$35)9

Each published research also carries its own actual cost in the frontmatter.

Can I keep research private?

Yes. Make your Atlas repo private (GitHub Pro enables private Pages), or keep Pages off entirely — you still get a themed, git-versioned archive.

What does a profile look like?

The optional profile.yml sits next to profile.example.yml in your Scout fork (gitignored). Every field is optional; sharpening only references a field when the topic naturally intersects it.

# used for "near me" and local rules
location: Ghent, Belgium

# ISO-639-1, ordered by preference
languages: [en, nl]

# ISO-4217 for any pricing
currency: EUR

# optional "(expert | proficient | learning)" hint
interests:
  - programming (expert)
  - AI (proficient)
  - table tennis
Does it work on GitHub-hosted runners?

Not out of the box. The self-hosted runner is what lets Scout use your Claude subscription instead of paying per-request. You can flip runs-on to ubuntu-latest and inject an API key, but you lose the cost advantage.

What data leaves my network?

Research prompts and responses go to Anthropic. Issue comments and published research go to GitHub. During research, Scout fetches URLs you'd otherwise open in a browser (WebFetch, WebSearch). Nothing else — in normal operation.

Is this sandboxed? What if a malicious page prompt-injects Claude?

It's not sandboxed. Scout runs Claude with --dangerously-skip-permissions inside a container with unrestricted network access. A malicious page reached during research could, in principle, instruct Claude to misuse the credentials on the runner or scan your LAN.

What's actually on the runner:

  • GITHUB_TOKEN — scoped to issues:write + contents:read on Scout only. Can spam/close issues, cannot push.
  • Atlas SSH deploy key — push access to your Atlas repo. Worst case: attacker publishes anything they like at your github.io URL.
  • Claude Code auth — your subscription. Attacker runs Claude on your bill.
  • Network position on your LAN — scan routers, NAS, Home Assistant, etc.

Egress allow-listing, capability drops, and a proper seccomp profile are on the roadmap. Until then: this is a YOLO box. Keep it on a network you trust and don't point Scout at topics likely to pull in hostile content.