Data and API for Researchers, Risk Teams, and Engineers

MCP Endpoint

The full data layer is exposed as an MCP server — Anthropic's open standard for letting language models call tools. Any MCP-compatible client can query the indicators and county scorecards directly: Claude Desktop, Claude Code, Cursor, plus any in-house agent built on the MCP SDK. Read-only, rate-limited, auth-stub-ready for higher tiers, and free to install.

Install

Download the MCP server source files, install the Python dependencies, and add the server to your client's config. Full setup notes live in the MCP server README.

{
  "mcpServers": {
    "american-default": {
      "command": "/absolute/path/to/venv/bin/python3",
      "args": ["-m", "scripts.machine_layer.mcp_server"],
      "cwd": "/absolute/path/to/american-default",
      "env": {
        "PYTHONPATH": "/absolute/path/to/american-default"
      }
    }
  }
}

Restart Claude Desktop. The six tools appear under the hammer icon. Other MCP clients (Cursor, in-house agents using the SDK) accept the same stdio invocation.

Tool surface

The five legacy tools carry schema_version: "v1" and retain their pinned response shapes. get_indicator_v2 requires response_version: "v2" and returns schema_version: "v2". Unsupported versions fail closed rather than falling back to a different shape.

Worked example — cross-correlation from inside Claude

A risk modeler asks Claude (with the MCP server installed) to find what follows credit-card delinquency in the fully validated scanner output. Claude invokes get_cross_correlations with the stored indicator slug and synthesizes the response.

get_cross_correlations(slug="the-late-fee")

{
  "indicator_slug": "the-late-fee",
  "as_leader": [
    {
      "leader_slug": "the-late-fee",
      "leader_name": "Delinquency Rate on Credit Card Loans",
      "follower_slug": "charge-off-rate-on-all-loans",
      "follower_name": "Charge-Off Rate on All Loans",
      "lag_quarters": 3,
      "correlation_r": 0.6565505852569093,
      "crises_validated": 2,
      "granger_p": 0.000632,
      "oos_validation_r": 0.8273
    }
  ],
  "as_follower": []
}

The same API workflow can be used with any of the 103 cataloged indicators. Cross-correlation results came out of a five-filter scanner (FDR-corrected cross-correlation, first-differenced series, multi-crisis validation, Granger testing, out-of-sample validation) that tested 87,296 pair-lag combinations and produced 9 fully validated relationships. Methodology lives at /methodology/; the systematic results return as JSON at /api/research/leading-indicators.json.

Source: scripts/machine_layer/. Probe mode at python3 -m scripts.machine_layer.mcp_server --probe emits a JSON handshake and exits — useful for CI smoke tests.

Indicator Catalog

103 indicators across 10 categories — debt stress, buffer depletion, financial conditions, labor market, cost pressure, plus AWI workforce indicators and demographic context. Every indicator has its own page (live data, chart, source attribution, Wire-voice editorial), full historical JSON, downloadable CSV, and machine-readable citation files.

Per-indicator URLs

Substitute any indicator slug (e.g. the-buffer, auto-loan-delinquency-90-days, cpi-inflation-rate-all-items) for {slug} below. The full slug list is in the human-readable indicator index, the machine-readable /api/indicators.json, or the /llms.txt that LLMs hit on first contact.

Browse the full indicator list at /indicators/. The /sources/ index groups indicators by federal publisher (FRED, BLS, NY Fed, Census, etc.) for users who think in publisher-units rather than topic-units.

Bulk Endpoints

When you want everything in one fetch — to load into a notebook, a backtest harness, or a model committee deck.

County-level scorecards are best fetched one at a time via the MCP get_county_scorecard(fips) tool — the JSON payload is ~2.5 KB per county and the rate limiter is generous enough to walk the full 3,144-county set inside an hour from a single client. The scorecard JSON includes the composite score, all five domain scores, national and state ranks, and pre-baked citations. Source: get_county_scorecard.py in the MCP server.

A county-scorecard PDF (single-page, model-committee-ready, branded) is published for every county at https://pdf.americandefault.org/scorecards/{fips}.pdf. Useful when a stakeholder wants the data on paper instead of in JSON.

Methodology

The full ADI methodology lives at /methodology/adi/ — five domains (Delinquency, Default & Legal, Debt Burden, Labor, Safety Net & Buffer), each the mean of its member indicators' Hazen percentiles within their own full quarterly history, equal domain weights, and national band labels that always publish with the literal reading. The published history holds the global financial crisis at its peak (92.3 in 2009-Q3) and the stimulus era at the trough (22.4 in 2021-Q4). The page includes the production scoring rules, the replication code link, and citation templates.

The County Distress Index methodology is published as a PDF at /methodology/cdi-methodology.pdf — five equal-weighted domains, the current admitted indicator set, distress-oriented county percentiles, and geography labels expressed as ranks plus score labels.

The systematic leading-indicator pipeline (cross-correlation + first differences + multi-crisis validation + Granger causality + out-of-sample) is documented inline at /methodology/#leading-indicators. The replication command is in the README at the repo root.

Citation

The license is permissive — Creative Commons Attribution 4.0. Use the data in commercial work, model committees, products, and journalism. Attribution is the only requirement. The canonical institutional name is American Default Research; the brand and URL is American Default; the composite is the American Distress Index (ADI); the county-level composite is the County Distress Index (CDI).

Templates for APA, MLA, Chicago, BibTeX, and news-copy short form are at /methodology/adi/#cite. Per-indicator BibTeX and RIS files are at /citations/{slug}.bib / .ris. The full canonical attribution block — including incorrect forms not to use — is at /llms.txt § Canonical Attribution.

FRED-derived series carry FRED's own attribution requirements (St. Louis Fed Terms of Use, including the no-AI-training clause). Downstream republishers should pass those through. Each indicator page lists its source attribution at the bottom; the per-indicator JSON and CSV downloads include it inline.

Talk to us about commercial use

Free with attribution covers virtually every use we have seen so far — research, journalism, product features that cite us, model-committee presentations. Contact us about anything that doesn't fit that shape: partnerships, embedded uses, white-label requests, custom data work, or a service-level arrangement for operational support.