stackscout
nothing gets in untested
More skills don't make an agent better. A clear line does — the right few, in the right order. We test for that, and publish everything that didn't hold up.
works with Claude Code · Cursor · any agent that speaks MCP
✕a pile of skills
Ten skills around it, none of them connected. No order — so it picks by chance.
✓a scouted set
Through the connector it gets a whole stack — the rest fades out and only the parts that fit stay connected.
What we actually do
We use it before you do.
Every skill ships in our own work first. Demo-only doesn't count.
Then we test them together.
A set has to hand over cleanly — or it isn't a stack.
The reasoning ships with it.
Skills without instructions are just files. So: order, guide, and what it can't do.
596k stars
GitHub stars of the 10 tools we point to — their number, not ours. the picks
6 turned down — with the reason.
read themAuthor, licence, check date — on every entry.
the libraryNo user counts, no download badges, no star ratings — we don't have honest ones yet.
What's inside
Same scouted library, three shapes. Pick by what you're missing right now — the first line on each card tells you which one that is.
Skills
You know exactly what you need.
.claude/skills/
├─ design-direction/SKILL.md
├─ impeccable/SKILL.md
└─ taste/SKILL.md
▸ “clean up this page”
✓ agent picks the right one — no setup
A small SKILL.md your agent reads on its own — with the author, the licence, and one line on what it can't do.
Browse skillsToolstacks
You know the goal, not the tools.
goal: a site that doesn't look AI-made
+ the guide that ties them together
→ one prompt sets up all of it
Skills, repos and guides that belong together — in the order they're meant to run. One prompt installs the set.
Browse toolstacksGuides
You want to decide for yourself.
→ read it, or let the agent pull it
How agents actually behave, measured on real work — including what we got wrong and what we'd skip next time.
Read the guidesUnderneath the stacks sit the repo picks — the tools we actually run.see the repo picks
How we scout
Four stages between “looks interesting” and a page on this site. Each one is allowed to kill the candidate — and most of them do.
We look where the work happens: real repos, real issues, what people actually run. Not top-10 posts that copy each other.
throws out: hype, dead projects
We ship something with it before we judge it. Anything that only shines in its own README dies at this stage.
throws out: demo-ware, half-finished
Author, licence, and the date we checked — on every card. Copyleft and non-commercial stay out, so you can earn money with what you build.
throws out: GPL/AGPL, NC, unclear weights
What's left gets a page, including one line on what it can't do. What isn't gets a rejection with the reason and the source.
keeps: the short list you see here
what we won't do
The proof of a filter is the pile next to it — every rejection is public, with the reason, the source and the date.
The connector
live · MCP over OAuthOne connector puts the whole scouted library inside your agent — every skill, stack, guide and repo pick. Everything we add later is simply there, in every project, without you touching a folder.
without it — every time
✕ five steps · per project · per update
with the connector — once
claude mcp add --transport http stackscout https://mcp.stackscout.app/api/mcp
what your agent can call after that
list_skills()
what's in the library right now
get_skill()
the full playbook for one skill
list_stacks()
kits by goal
get_stack()
the whole kit + setup prompt
search()
one query across everything
list_repo_picks()
the tools we actually run
It calls them by itself when a task needs them — you just work.
7 days free · no card · connected in about a minute · cancel any time
Prefer files? Every free skill is still a one-line download.
Scout with us
Send it over — a skill, a repo, a guide we should write, or something on this site that's plain wrong. The good ones go into the next scout. No account, no email needed.