Scout Dog — Discovery & Monitoring para Claude Code
apswin/scout-dog
Busca, filtra e resume conteúdo web por comando em português natural
Ver no GitHubSkill que automatiza buscas na web e em sites específicos a partir de descrições em linguagem natural. Filtra, lê e resume resultados em notas estruturadas, sem depender de APIs externas — usa apenas a sessão Claude Code. Ideal para pesquisa de conteúdo, monitoramento de listagens e descoberta de fontes.
Instalação
npx convez-ai apswin/scout-dog1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Pesquisar threads Reddit, artigos ou perfis X por tema específico
- Monitorar listagens (imóveis, empregos, produtos) com filtros customizados
- Descobrir recursos, ferramentas e referências para projetos de pesquisa
- Acompanhar tendências de conteúdo em horizontes de tempo específicos
Veredito

Scout Dog resolve bem o caso de pesquisa semi-automática com filtros inteligentes, mantendo a UX conversacional. Força prática: não exige chaves de API externas. Limitação: depende de acesso web aberto e da capacidade do Claude ler páginas dinâmicas (Playwright é opcional).
README
A tradução para o português deste README ainda está na fila.
scout-dog (for Claude Code)
A discovery skill for Claude Code. Describe what
you're hunting in plain English and scout-dog runs boolean web searches, filters
the hits against your KEEP/DROP criteria, reads the pages it can reach, and writes
a curated findings note to ./runs/.
It finds, reads, and shortlists — it never does outreach or contacts anyone.
"find Reddit threads naming good SEO audit tools"
"find 50 X profiles who post about GEO and AI search"
"round up recent articles on agentic RAG"
"new SF 2BR apartments under $6500/mo on Craigslist, last 24h"
│
▼
gather → dedupe → filter → fetch & summarize → ./runs/findings_<date>.md
Two gather modes, picked automatically from your ask:
- web — boolean search across the open web (the default).
- site — pull one site's own filtered listings/feed, with structured filters (price, beds, location) and recency ("last 24h"). Uses the site's RSS where available; for anti-bot sites it reads the results page via Claude's fetch.
How it works
One intelligence layer: the Claude model you're already chatting with does the briefing, the KEEP/DROP filtering, and the summarizing/classifying. Plain-Python scripts do only the deterministic mechanical work (search, dedupe, fetch, render) and emit JSON. There's no second model and no API key — page-heavy work is fanned out to Claude Code subagents so it never floods your main chat context.
See SKILL.md for the full pipeline and
references/brief-schema.md
for every knob.
Install
Option A — use it inside this repo (simplest). Clone, open Claude Code in the repo, and just ask:
git clone https://github.com/apswin/scout-dog
cd scout-dog
claude # then: "find Reddit threads about ..."
The skill in .claude/skills/scout-dog/ is auto-discovered.
Option B — install it globally for use in any project:
cp -r .claude/skills/scout-dog ~/.claude/skills/
First run
The first time you invoke it, scout-dog runs a one-time setup: it creates a
virtualenv, installs the base dependencies (ddgs, requests, beautifulsoup4),
asks once whether you want the optional Playwright browser tier for
JavaScript-heavy sites, and runs a smoke-test. Every run after that skips straight
to hunting.
You can also run setup manually:
bash .claude/skills/scout-dog/scripts/setup.sh # base
bash .claude/skills/scout-dog/scripts/setup.sh --with-playwright # + browser
Requires Python 3.9+.
Output
Each run lands in its own folder under ./runs/ (gitignored):
runs/<date>-<name>-<id>/
├── brief.json the contract you agreed to
├── 01_raw.json 02_deduped.json 03_findings.json
├── findings_<date>.md ← the deliverable you read
├── findings_<date>.csv
└── _run.md the funnel (searched → kept → fetched → snippet-only)
Give a hunt a slug to turn it into a re-runnable monitor: its runs accumulate
under runs/<slug>/ and a _seen_urls.txt registry skips anything already found.
Scheduling
Scheduling is optional. After a successful run the skill asks whether you want it to
run automatically and show only new results each time. If you say yes, it sets up a
recurring run (daily for fast-moving listings, weekly for content discovery) at a
time you choose, defaulting to your local timezone. Scheduling is driven by Claude
Code's own scheduler — not cron on the scripts — because the filtering and
summarizing need Claude. You can also just ask "run the <slug> monitor" any time.
Honest limits
Login-walled sites (X, LinkedIn, Instagram, …) can't be read without auth, so for those scout-dog returns the search snippet and says so. The optional Playwright tier helps with JavaScript-rendered pages, not with login walls.
License
MIT — see LICENSE.