10 read-only tools over the same records the pages publish
An MCP server, a JSON API, RSS feeds, dataset downloads, and a Markdown version of every page.
Instructions for agents and the people who configure them. Every record returned includes data_source, source_url, authority_tier and is_estimated, the same four fields the pages render. Access is free and requires no login.
Machine-readable access points
- MCP server
https://scrutica.com/api/mcp: 10 read-only tools, reference resources and ready-made analysis prompts over the site’s records. The LLM-optimized reference is the agent guide; the discovery card sits at /.well-known/mcp/server-card.json. Listed in the MCP Registry asio.github.davidgringras/scrutica.- Markdown twins
- Append
.mdto any page URL for a Markdown rendering (for example /methodology.md), per the llms.txt suffix convention. The site index is /llms.txt; the complete single-file reference is /llms-full.txt, rebuilt from the same records before every deploy, with a check that fails the build if the committed file and its generator have come apart. - Feeds
- RSS for Entity List changes (whole log, or per organization), sovereign-program evidence (all programs, or per program), corrections (every public disclosure of a value we got wrong), and the Record (daily digests of records added and updated). A JSON audit feed reports BIS cross-reference cascade deltas at
/api/export-controls/cascade-feed. - Dataset downloads
- Packaged bundles with CSV, schema, data dictionary, SHA-256 checksum and Croissant JSON-LD, from Data & Citation. Each bundle page reports measured column completeness, so a client can tell before downloading whether a column is worth parsing. Citation forms are at the citation guide.
- JSON API
- Public read endpoints, described in OpenAPI 3.1 at /api/openapi.json. Content negotiation is live: request any page URL with
Accept: text/markdownand the Markdown twin comes back at the page’s own URL.
The tools
scrutica_searchFree-text search over facilities, companies, and sovereign programs; returns canonical ids with per-row source attribution.scrutica_get_facilityComplete facility record by canonical id: capacity, operator and owner, energy profile, campus relations, source attribution, estimate flags, and a ready-made citation.scrutica_get_companyOrganization record by canonical id: type, jurisdiction, identifiers, and ownership references.scrutica_query_export_controlsBIS Entity List designations, each anchored to its Federal Register notice (tier-1 sourcing).scrutica_entity_list_changesWhat changed in the BIS Entity List: per-notice change events, removal actions, and a weekly activity rollup.scrutica_get_supply_chainSupplier–customer edges for one or more organizations, weighted by supply share and 3-month price correlation.scrutica_get_sovereign_programNational sovereign-AI program records: announced vs disbursed capital, reality ratio (disbursed over the government-only announced figure, which is returned alongside it), status, and dependency structure.scrutica_get_scenarioThe evidence base behind a geopolitical compute-risk scenario — the disruption cases it models and the facilities, fabs, cables or weapon systems they reach, each item carrying its own source and authority tier, and each capped list its true total — plus the URL of the interactive propagation analysis, which stays interactive-only.scrutica_estimate_flopsPeak and sustained FLOP estimates for a hardware configuration, each returned with the estimation bounds around its point estimate and is_estimated set.scrutica_get_methodologyCanonical methodology summary and URL for a Scrutica metric: how a given number was calculated.
The server also exposes the agent guide, the full site reference and per-topic methodology summaries as MCP resources. Three prompts come ready-made, each carrying its citations through to the answer: facility_brief writes up one facility with every figure sourced, export_control_changes_brief reports Entity List additions and removals, and verify_scrutica_number walks the sources behind a single figure.
Connect the MCP server
Transport is streamable HTTP. No authentication. Rate limit 60 requests per minute per address, with Retry-After on a 429. A plain GET on the endpoint returns this same connection information as JSON.
- claude.ai and Claude Desktop
- Settings → Connectors → Add custom connector, then paste
https://scrutica.com/api/mcp. - Claude Code
claude mcp add --transport http scrutica https://scrutica.com/api/mcp
- Cursor
- One-click install (Cursor’s documented deeplink), or add to
.cursor/mcp.json:{ "mcpServers": { "scrutica": { "url": "https://scrutica.com/api/mcp", "transport": "streamable-http" } } } - VS Code
- One-click install (the documented
vscode:mcp/installscheme), or use the same JSON server entry as Cursor.
The data contract
What a response means, so it can be relayed without distortion.
authority_tier- 1: primary measurement or regulatory filing (SEC EDGAR, Federal Register). 2: research database or curated corpus aggregating primary sources. 3: press or analyst secondary. 4: Scrutica-derived, with documented methodology and adjustable parameters.
is_estimated- True where the number was derived rather than measured, and a derived number carries its bounds where the method produces them. A null anywhere in a response is not a zero: the source published no value, and filling it in invents one.
- Campus records
- Several facility rows can describe one physical site at different scopes. Responses mark the relation (
campus_id, canonical-for-power member); summing member megawatts double-counts. - Corrections
- Where the site showed a wrong value, the disclosure is public and permanent at the correction record. Routine additions land in the Record, which reports its ingest cadence so a client can see whether a quiet feed means a quiet week or a stalled pipeline.
- Freshness
- Ingest crons refresh sources on daily-to-weekly cadences; the data snapshot behind rendered counts is dated (currently 2026-09-05). Entity records have their own update timestamps, and sitemap
lastmodvalues reflect source updates rather than request time.
Crawling and reuse
AI crawlers, retrieval fetchers and search indexers are explicitly welcome. robots.txt allow-lists the documented agent families by name and states affirmative consent for training, retrieval and search use; only admin paths are disallowed. Identify your client with a real User-Agent and contact URL, and respect the published rate limits.
Reuse licence: CC BY-SA 4.0 for the site’s own rendered data, except where a surface states otherwise. The packaged dataset bundles are CC-BY-4.0, stated on their own pages and in every citation form. Source datasets originating in licensed corpora are stored as extracted relational facts with provenance preserved, and the raw licensed datasets are not redistributed.
When relaying a number, keep its source_url and its tier with it, so whoever reads your answer can check the filing directly. Entity pages have a Cite panel, and MCP facility and company responses include a ready-made cite string with the canonical URL and access date.