AI agents — install + integration guide

Wire Claude / ChatGPT / Perplexity / Gemini / Microsoft Copilot / Cursor / Continue / Windsurf agents to Policy Window's primary-source-cited AI-governance catalog. Read-only retrieve-and-cite per charter §7.1.b (no submit-on-behalf-of-user). Open-standards first; every path here survives any single AI platform shutting down.

1 · MCP descriptor (recommended; works today)

Policy Window publishes a Model Context Protocol descriptor at https://policywindow.org/mcp.json documenting the 8 read-only tools (search_catalog, get_instrument, get_topic_coverage, get_coverage_cell, compare_jurisdictions, list_changes, cite_article, glossary_lookup), the response-envelope shape, the charter-compliance constraints, and the public endpoints to wire them against.

Until the full @policywindow/mcp-server npm package ships (on the roadmap, on the roadmap (~weeks 4-9; see /wiki/roadmap)), agents wire the tool surface directly against the existing public endpoints documented in the descriptor.

# Fetch the descriptor + see the tool surface
curl -s https://policywindow.org/mcp.json | jq '.tools[].name'

2 · llms.txt + sub-audience variants (auto-discovered)

AI assistants (Claude, ChatGPT, Perplexity, Google AI Overview, Gemini) increasingly look for /llms.txt at the site root as a curated guide to high-value content. Policy Window publishes:

/llms.txt — canonical short-form index
/llms-full.txt — full-context companion (methodology + charter + AI-disclosure inlined)
/llms-researchers.txt — for academic-researcher agent contexts
/llms-journalists.txt — for journalist-on-deadline agent contexts
/llms-procurement.txt — for procurement-evaluator agent contexts
/llms-regulators.txt — for regulator-staffer agent contexts
/llms-advocates.txt — for civil-society-advocate agent contexts
/llms-developers.txt — for developer-integration agent contexts

Configure a Claude Project / Custom GPT / Gemini Gem to fetch the appropriate sub-audience variant as its system context; the agent will navigate the catalog with the right priorities for its workflow.

5 pre-built platform templates → /wiki/templates ships forkable setups with the system prompt and knowledge files wired up. CC0 1.0 — fork freely.

Claude Project — system prompt + uploaded knowledge files for Claude.ai Projects.
Custom GPT — OpenAI GPT Builder configuration with action schema for /api endpoints.
Gemini Gem — Google Gemini Gem instructions and starter prompts.
NotebookLM — Google NotebookLM source bundle with all llms-*.txt variants pre-loaded.
Microsoft Copilot — Copilot Studio + custom-instructions setup referencing the public OpenAPI surface.

3 · Schema.org structured data (Google AI Overview eligibility)

Every PW article emits schema.org JSON-LD (Article + Legislation + BreadcrumbList + FAQPage). The hub at https://policywindow.org/wiki adds a Dataset node. Every topic article emits per-cell ClaimReview nodes (the single highest-leverage AI-Overview citation move; peer platforms with correct ClaimReview report 48%+ AI-Overview citation rate).

4 · Open APIs (no auth, CORS-open)

GET /wiki/catalog/json — full structured catalog (~3MB, CC0 citation graph)
GET /api/search?q=foo&kind=instrument&limit=50 — ranked search (10KB; resolve natural-language refs to slugs)
GET /api/openapi — OpenAPI 3.1 spec
GET /wiki/[slug]/datacite.xml — DataCite 4.4 per article
GET /wiki/changelog/feed — RSS 2.0
GET /mcp.json— MCP descriptor (this page's primary integration target)

5 · Standard response envelope

Every tool exposed by @policywindow/mcp-server, the citation routes (/wiki/citing-us/citations.json), and the per-audience /llms-*.txt variants return the same envelope. Agents handling Policy Window responses should treat this as a stable contract:

{
  "data":          /* tool-specific payload */,
  "citation": {
    "wikiUrl":       "https://policywindow.org/wiki/<slug>",
    "primarySource": "<official-source-URL>",
    "confidence":    "high" | "medium" | "low" | "pending"
  },
  "attribution":   "Source: Policy Window AI Governance Catalog (https://policywindow.org/wiki). Snapshot: ?asOf=YYYY-MM-DD. CC0 1.0 Universal.",
  "asOf":          "YYYY-MM-DD",
  "schemaVersion": "0.1"
}

citation.wikiUrl — always points at the pinned ?asOf= version so the agent's citation is snapshot-stable.
citation.confidence — per-claim confidence tier; agents must surface it when citing PW (see charter §7.6 reproduction below).
attribution — pre-formatted CC0 attribution string; agents must reproduce verbatim under rendered answers.
schemaVersion — breaking-change marker. Consumers handling envelopes should accept 0.x today and gate on a major bump before parsing breaking-change fields.

6 · MCP server install (npm package, alpha)

The npm-publishable source lives in the monorepo at packages/mcp-server/ (see the README). Once the SDK dependency is installed and the package publishes, install globally:

npm install -g @policywindow/mcp-server

Wire it into Claude Desktop by adding the following to your claude_desktop_config.json:

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

Caveat: the @modelcontextprotocol/sdk dependency is declared in the package.json but not yet npm-installed in the monorepo (alpha release pending). Until then, wire the descriptor at /mcp.json directly against the existing public endpoints (section 1) — the tool surface + envelope contract are stable; only the npm distribution awaits the SDK bump.

7 · Roadmap (later in 2026; see /wiki/roadmap)

@policywindow/mcp-server npm package (on the roadmap, ~2026-Q3; see /wiki/roadmap) — stateless stdio + HTTP transport with the 8 tools wired
/wiki/catalog/cells/[instrument]/[topic] — per-cell JSON anchor
/wiki/catalog/rdf — Turtle export with owl:sameAs to Wikidata + ELI + CELEX
/wiki/sparql.json — live SPARQL 1.1 (SELECT/ASK) over the catalog RDF graph; console at /wiki/sparql
Provision-level verbatim text backfill (525 cells × 6 min editorial via Coverage Games quarterly cycles)
/wiki/catalog/coci.csv — OpenCitations-compatible outbound citation dump
/embed/cell/[instrument]/[topic] + oEmbed JSON for CMS-pasteable citation cards

Charter constraints for AI-agent integrations

Charter §7.6 — No covert persuasion or political microtargeting (verbatim):

“Policy Window does not deploy AI for influence operations. All editorial output is signed (by the editorial board for catalog content; by the expert validation bench for research-module content) and dated. No anonymous content, no per-audience-tailored persuasion, no political microtargeting, no covert-influence operations. This subsumes the existing §7 microtargeting bright line and clarifies that it applies to every product layer (free catalog, paid exports, paid research modules).”

Charter §7.7 — No unapproved human-subject research (verbatim):

“The on-demand expert validation bench conducts expert and stakeholder interviews when a customer engagement requires it. Every such interview requires: (1) informed consent in writing from each participant, (2) declared data retention (default: structured summary retained; raw recordings deleted after the validation cycle), (3) anonymisation by default (named attribution only with explicit per-instance consent), (4) a written protocol shared with the participant before the interview, and (5) a named expert researcher accountable for the protocol. No covert observational research; no human-subject work without an approved protocol; no synthetic-respondent substitution.”

What this means for agents: agents using the Policy Window catalog must (a) reproduce the attribution envelope field verbatim so the source is visible to end users, (b) use ?asOf= pinning so a citation can't silently change beneath a reader, and (c) defer legal advice — the catalog is primary-source citation infrastructure, not jurisdiction-specific counsel.

Read-only retrieve-and-cite — no write tools, no submit-comment, no post-to-regulator, no subscribe-to-alerts. Per charter §7 + §7.1.b (no client lobbying); MCP cannot take actions on behalf of users.
No personalisation — every response is identical given identical input (§7.6 no microtargeting).
No prose generation by PW — the server returns structured catalog data + verbatim primary-source quotes; agents author their own prose. Confidence tier + citation envelope constrain hallucination.
Confidence tier always surfaced — every response includes citation.confidence (high/medium/low/pending); agents must surface the tier when citing PW.
Permanent ?asOf= URLs— every returned wiki URL is pre-built with today's date so the agent's citation is permanently snapshot-pinned.
Attribution preservation — every response includes an attribution envelope field; agents must reproduce it verbatim under rendered answers.