agenticschema: Skill for Claude Code

Agentic Live Search (@agenticschema/server)

Never dump raw HTML (143,770 tokens/page mean) or raw Readability/Markdown prose (2,752 tokens/page mean) into your context window when researching the web. Use AgenticSchema to extract verified epistemic evidence at 6.7× to 28× lower token cost (413 tokens at L1, 110 tokens at Submodular budget=120, 99 tokens at L0), with explicit cross-representation consistency, temporal validity, per-facet sufficiency diagnostics, and an Expected Value of Information (EVI) controller.

1. Preferred Mode: Dynamic Live Search MCP Server (--live)

If your host supports MCP servers, register @agenticschema/server in --live mode:

{
  "mcpServers": {
    "agenticschema-live": {
      "command": "npx",
      "args": ["-y", "@agenticschema/server", "--live"]
    }
  }
}

This exposes three dynamic live-search tools that accept arbitrary public http(s) URLs on the fly (with built-in SSRF protection):

1. fetch_evidence({ url, query?, resolution?, tokenBudget? })
   • Fetches a single URL and distills it into an epistemic context block.
   • Start with resolution: "L1" (default, 413 tokens mean) or set query + tokenBudget: 120 (110 tokens mean) to activate the Multi-Facet Submodular Query Bottleneck.
   • Inspect the ⚖️ Epistemic Sufficiency and 🎯 Controller lines in the output header:
     • ANSWERABLE / STOP_AND_ANSWER → Every required query facet satisfies $\text{Evidence}(f_i) \ge \tau_i$ (SATISFIED). Answer directly and cite the inline sigils.
     • PARTIAL_ESCALATE + ESCALATE_L2 → One or more facets are NOT_PUBLISHED in structured markup; re-call fetch_evidence on the same URL with resolution: "L2".
     • PARTIAL_ESCALATE + QUERY_MACHINE_TWIN or FOLLOW_FRONTIER → Follow the target same-origin JSON twin or high-scent frontier hop (e.g. .pdf datasheet).
     • ABSTAIN → Explicitly report which facets are NOT_PUBLISHED, NOT_FOUND, CONFLICTED, or UNVERIFIABLE instead of guessing from parametric memory.