emet tool reference

Use this doc when you already know you want emet, but need the shortest path to the right call shape.

Pinned: README · quickstarts · examples · pipeline

Public tools

Tool	Use it for
emet	live research, docs lookup, comparisons, changelogs, security checks, papers, local file grounding, community/media retrieval, checkpointed sessions
web_fetch	raw extracted text for one known URL

Pick the right mode

Mode	Best for
fast	current facts, quick doc checks, authority-first lookups
deep	comparisons, ambiguous questions, broad research, community/media work
code	docs, APIs, repos, READMEs, code-oriented answers
academic	papers, DOI/arXiv metadata, scholarly sources

High-value options

Trust and source control

• requireAuthoritative: true — force stronger sources for factual/high-risk work
• sourcePolicy: "official-only" — bias toward official docs/pages
• preferRecent: true — better for current-product or release questions
• allowedSources: [...] — soft source hints for ranking; if every entry is a concrete host or host/path, they are also applied fail-closed
• hostAllowlist: [...] — strict fail-closed host or host/path filter when you already know the canonical sources

Raw source access

• web_fetch({ url }) — best when you already have the URL; faster than deep mode for targeted page reads
• options.selectedUrls: ["https://..."] — fetch and summarize known URLs through emet; works with or without an interactive session
• options.rawPages: true — best when you want a normal emet answer plus full source text in pageTexts[]
• options.requirePrimarySource: true — require official/primary evidence where possible; aliases to stricter authority guardrails

Local repo grounding

• options.files: ["package.json", "README.md"] — merge local evidence into the research run

Community/media retrieval

Use options.platforms explicitly for read-only community/media sources:

• hn
• reddit
• v2ex
• github
• rss
• youtube

Important: community/media results are signals, not automatic truth. For CVEs, outages, deprecations, vendor status, legal/medical topics, and similar high-risk claims, follow with authoritative sources.

Checkpointed sessions

• interactive: true — return compact state and next choices; today this is most useful with explicit community/media platforms
• sessionId — continue a bounded session
• action: "search" | "refine" | "fetch" | "synthesize" — continue the session intentionally
• queryOverride — refine the question
• selectedResultIds — fetch selected results from a checkpoint response
• deepResearchConfig.depth — follow-up depth, 1-3
• deepResearchConfig.breadth — queries per depth layer, 2-4
• deepResearchConfig.concurrency — concurrent searches/fetches, 1-4

interactive: true is for checkpointing. platforms is for community/media retrieval. They are related, but not the same thing. If you already know the URL, use web_fetch instead of an interactive round.

Good call patterns

Current docs

{
  "query": "current MCP sampling docs",
  "mode": "code",
  "options": {
    "requireAuthoritative": true,
    "hostAllowlist": ["modelcontextprotocol.io", "github.com/modelcontextprotocol"]
  }
}

Security / CVE verification

{
  "query": "Verify CVE-2026-12345 from vendor and NVD sources",
  "mode": "deep",
  "options": { "requireAuthoritative": true, "preferRecent": true }
}

Community + GitHub checkpoint

{
  "query": "What are developers saying about React 19 upgrade pain?",
  "mode": "deep",
  "options": {
    "platforms": ["hn", "github"],
    "interactive": true,
    "maxResultsPerPlatform": 5
  }
}

Raw text after a normal research run

{
  "query": "MCP tool schema docs",
  "mode": "code",
  "options": { "rawPages": true }
}

Anti-patterns

• Do not use interactive: true for every call. Use it only when you want checkpoint/next-step control.
• Do not treat community results as enough for high-risk factual claims.
• Do not rely on allowedSources alone for strict verification; use hostAllowlist when source boundaries must fail closed.
• Do not use rawPages: true by default; it makes responses much larger.
• Do not use deep when fast, code, or web_fetch is enough.
• Do not use web_fetch as a search tool; it expects a known URL.