Search Sh

Search.sh — Web Search for Agents

Search.sh is a search engine built for AI agents. Use the web_search and web_extract tools to query the web and get structured results.

When to Use

Use web_search instead of browser_navigate for:

• Research tasks (market research, competitor analysis, trend tracking)
• Fact-checking and verification
• Finding current information (prices, news, releases)
• Any task that starts with "find", "research", "look up", "what is"

Use browser_navigate only when you need to:

• Interact with a website (click, type, login)
• Take screenshots
• Access authenticated content
• Post content to platforms

Tools

web_search

Search the web and get an AI-synthesized answer with sources.

web_search query="best vector databases 2026" mode="fast"
web_search query="comparison of Next.js vs Remix for production apps" mode="deep"

Modes:

• fast (default) — single search + AI answer. 3-8 seconds. Use for quick lookups.
• deep — generates sub-queries, searches in parallel, extracts full page content. 15-30 seconds. Use for thorough research.

Returns: answer, sources (title + URL + snippet + published_at / modified_at / date_confidence), citations with attribution, conflicts (sources that disagree across time, newer first), confidence score (0-1), related queries.

Dates and recency:

• since="2026-01-15" — only pages the engine dates on/after that day. recency="past_month" (past_day|past_week|past_month|past_year) — the same, relative. freshness_boost=true — newer pages ranked higher, older ones kept.
• Check the dates before trusting a fact: a modified_at whose date_sources.modified_at is http:last-modified is the site's deploy time, not an update.
• When conflicts is non-empty, do not average the two claims: prefer the newer side when dated is true, mention the date in anything you pass on, and re-query with since set just before the older source's date if you need corroboration.

web_search query="studios powering <brand> lobby" since="2026-03-01" freshness_boost=true

web_extract

Extract clean text from URLs. Use after web_search to read full content from specific sources.

web_extract urls=["https://example.com/article"]

Returns cleaned text (scripts/nav/footer removed), max 5000 chars per page, plus the page's own date (published_at, modified_at, date_confidence, date_sources). Up to 10 URLs per call.

Research Pattern

For thorough research:

1. web_search with mode="deep" to get an overview + sources
2. web_extract on the most relevant source URLs for full content
3. knowledge_write to save findings for future reference

Setup

Requires a Search.sh API key:

vault_set key=search_sh_api_key value=sk-sh_your_key_here

Get your key at https://search.sh/dashboard. Prepaid credits: fast search = $0.01, deep search = $0.05, extract = $0.01/URL.

Verify

• The intended other agent / tool / channel actually received the message; an ack, message ID, or response payload is captured
• Identity, scopes, and permissions used by the call were the minimum required; over-permissioned tokens are called out
• Failure handling was exercised: at least one retry/timeout/permission-denied path is shown to behave as designed
• Hand-off context passed to the next actor is complete enough that the receiver could act without a follow-up question
• Any state mutated (config, memory, queue, file) is listed with before/after values, not just 'updated'
• Sensitive material (keys, tokens, PII) was redacted from logs/transcripts shared in the verification evidence