Scout — lean web search + reader for AI agents
Free-first, provider-agnostic web search and page reader for AI agents. Token-efficient by design: highlights-first excerpts, reranked snippets, token budgets. Flat string/integer/boolean params so DSH, MCP, or OpenAI-style function calling can all use it. No API key needed.
Scout 0.5 is a lightweight evidence retrieval layer for agents, not a search engine. Scout exposes two tools around six retrieval capabilities: SEARCH candidate sources · READ evidence · FIND exact passages · follow LINKS to deeper sources · inspect PROVENANCE · support agent-led VERIFICATION. No internal LLM, no summarizer — cheap deterministic plumbing plus agent-facing affordances. Zero npm dependencies, plain fetch() + string parsing, no tracking.
Why Scout?
Many AI agents have reasoning and tool-calling capabilities but no usable web access in every deployment. Scout provides a lightweight, keyless web retrieval layer so the agent can decide when and how to gather external information. Scout does not plan research, summarize findings, or make decisions for the model.
Agent decides what to do.
Scout retrieves the information.
Agent reasons over it.
What you get
| Tool | What it does |
|---|
scout_search | Keyless-first web search (DDG + Google News RSS + Bing RSS + Wikipedia/HN/StackExchange/Reddit specials, optional Brave key or self-hosted SearXNG). site: support, alternatives expansion, provenance + health on every call. |
scout_read | Clean page text with text / highlights (~1/5 tokens) / links / find modes, plus Reddit, StackOverflow, HN, Discourse, PDF, and article fast-paths. |
web providers scout | Also backs the built-in web_search / web_fetch tools, so they work with no key configured. |
Details: agent usage · providers · architecture.
Quick start
scout_search with maxResults: 3-5 → compact [SRn] hits with dates, domains, and source types. Add alternatives (up to 4) only when the query is ambiguous or the first ranking missed.scout_read { url, query, view: 'highlights' } → extractive excerpts. Follow-ups on the same page are cached 10 min and free; long pages page with offset=N.- Hard question loop (agent-led): search → read 1–2 hits →
find the passage → follow links deeper → re-search if needed. Scout never runs this loop itself.
Full workflow, config, and output contracts: docs/agent-usage.md.
Health contract
Empty results and provider failure are different states. No results found means there may be no result; Partial results: / explicit outage means search infrastructure is degraded. Provider outages surface as a footer naming the failed backends, never as silent thin results.
Providers in brief
Works keyless out of the box. Optional BRAVE_API_KEY slots in first for better relevance; Google CSE is legacy-only (closed to new customers, off by 2027-01-01); optional self-hosted SearXNG slots ahead of everything. Full ladder and honest backend limits: docs/providers.md.
Honest limits (teaser)
- Keyless endpoints are undocumented and can change; every backend is failure-isolated so one outage never blanks a search.
- Bot-walled official docs are the hardest pages to read (direct 403 + rate-limited fallback) — the known ceiling, not a bug.
- No JS rendering; English-first; token counts are
chars/4 estimates.
Security model (SSRF guard, Jina privacy, fail-closed signed URLs): docs/security.md.
Install (this profile)
Already installed as a linked bundle (scout-dsh). After any code change:
- Restart the DSH app (the host requires a restart to load plugin updates — the installer reports
restart-required).
- Verify:
mount-status.json (next to index.js) appears after boot, e.g.
{"build": "0.5.0", "searchProvider": "registered", "tools": {"scout_search": "registered-via-…", …}}.
- Tools
scout_search / scout_read are now available to agents, and web_search keeps working even with no ANYSEARCH_API_KEY.
Fresh install elsewhere:
# from the plugin folder
dsh plugin add ./Scout-dsh
# or: DSH GUI → Settings → Plugins → Install → pick the folder
Or raw npm-style (after publishing):
npm install ./Scout-dsh
Then restart dsh web (or reload the profile). Tools scout_search / scout_read appear, and web_search keeps working even with no ANYSEARCH_API_KEY.
Docs
| Page | What's there |
|---|
| docs/architecture.md | Pipeline diagram, fusion/rerank, caches, circuit breaker, retrieval modes |
| docs/providers.md | Free-first ladder, Brave/CSE/SearXNG, intent routing, backend limits |
| docs/agent-usage.md | Full config, output contracts, lean workflow, forum/PDF reading |
| docs/benchmarks.md | Test gates, regression vs agent benchmark, pilot results |
| docs/security.md | SSRF guard, fallback privacy, honest limits |
Test
npm test # offline quality gates (L1–L5), all must-pass
Suites, frozen ranking bench, and the pilot agent benchmark: docs/benchmarks.md. The gates are regression guards, not proof.
Made with ✨ Muse Spark 1.3 Contributor ✨ — lean, transparent, and free for any agent.
