LLMS.TXT · MCP · BENCHMARK

Built for the agents your team ships, too.

Most of the traffic evaluating a synthetic-data vendor today is still a person with a browser tab. Some of it is already an agent — a coding assistant scaffolding test fixtures, an AI advisor being graded on whether it gets a planning answer right. We built a discovery and evaluation layer for that second kind of visitor, not as an add-on to the marketing site, but as a first-class surface: a plaintext manifest, a callable MCP server, and a benchmark with a real answer key.

Connect the MCP server View the Planning Benchmark

8 MCP tools50 benchmark tasksRead-only · free · no auth

FOUR SURFACES, ONE PRINCIPLE

Fetchable, not scrapeable.

Everything below is free, read-only, and returns structured JSON (or plain text for llms.txt) — no login, no rate-limit surprises, no HTML to parse.

llms.txt

A plaintext discovery file for coding agents (Cursor, Claude Code, etc.) building wealthtech features — points straight at the catalog, the manifest, and per-archetype samples.

View llms.txt

MCP server

One Streamable HTTP endpoint (JSON-RPC 2.0), 8 tools, no auth, no session state. Read-only by design — it's a discovery layer, not a metered API.

View the MCP endpoint

Agent manifest

The full archetype and bundle taxonomy in one JSON call, with a sample-household URL wired to every archetype — built for agents that would rather fetch than scrape.

View the manifest

Planning Benchmark v2

50 free evaluation tasks for AI financial-advice systems, answer keys by construction: required figures cited to primary sources, stale values enumerated with reason codes. Blind mode withholds the keys; the scoreboard publishes the protocol.

View the benchmark

DecisionSynth decision memory

Advisor-decision episodes with known-answer QA tasks — evaluate an agent-memory pipeline on recall, rationale, precedent, ordering, and rule attribution, or seed an agent's memory with realistic decision history.

Explore DecisionSynth

AI Eval Sets

Vintaged commercial eval packs behind the benchmark: adversarial threshold flips, decision-recall tasks, forbidden-figure tables, and a mechanical scorer — measure a system's Stale Figure Rate and gate deployments on it.

Explore AI Eval Sets

CONNECT

One endpoint, eight tools.

Point Claude, Cursor, or your own agent at a single MCP URL. No API key, no session setup — the server is stateless Streamable HTTP over JSON-RPC 2.0.

list_archetypesAll 71 synthetic household archetypes — id, name, life stage, wealth tier, household type.

list_bundlesProduct bundles with use cases and one-time-purchase links.

get_sample_householdOne deliberately-insufficient sample household per archetype — proves fidelity, not a usable corpus.

describe_household_schemaThe top-level fields of a synthetic household record.

get_purchase_linkThe purchase URL for a given bundle.

list_reference_parametersThe 2026 U.S. federal financial-planning parameters, optionally filtered by domain.

get_reference_parameterA source-verified parameter by key, with its primary-source citation.

get_planning_benchmarkPlanning Benchmark v2 tasks with forbidden-figure answer keys — pass blind=true to omit the keys.

~/.config · mcp.json

// add the server once — Claude, Cursor, or your app
{
  "mcpServers": {
    "wealthschema": {
      "url": "https://www.wealthschema.com/api/mcp"
    }
  }
}

describe_household_schema → result

{
  "persona": "household_type, state_of_residence, education, ...",
  "cash_flow": "gross/net income, expense breakdown, savings_rate_pct, ...",
  "assets": "checking, brokerage, retirement, real_estate, ...",
  "longitudinal": "monthly[] 96-month trajectory + methodology",
  // ...9 more fields, abridged for this page
}

EVALUATION, NOT JUST DATA

Can your advisor agent actually get it right?

14 tasks across 12 categories — net worth, savings rate, RMDs, IRMAA tiers, marginal brackets, estate exemptions, state-level rules. Each task gives a household's facts and a planning question with one deterministic, source-grounded answer. Most require looking up a live 2026 reference parameter and applying it correctly — the same failure mode that makes a plausible-sounding advisor agent wrong in production.

net_worthsavings_rateemergency_funddebt_to_incomeretirement_contributionrmdmarginal_tax_bracketirmaaestatehsastate_top_ratestate_retirement_taxation

Pass ?withhold_answers=true to fetch tasks without the answer key for blind evaluation — score the model's output yourself against the full benchmark.

v1 is superseded by /api/benchmark/v2 — 50 tasks with forbidden-figure answer keys and reason codes. The scored results and protocol live on the benchmark scoreboard.

task nw-001 · net_worth

{
  "question": "What is this household's net worth?",
  "household": {
    "filing_status": "single",
    "total_assets_usd": 850000,
    "total_liabilities_usd": 320000
  },
  "expected": { "value": 530000, "unit": "USD" },
  "scoring": { "method": "exact" }
}

WHY BUILD THIS AT ALL

An agent is a visitor, not an edge case.

The full validated corpus is still a one-time purchase, delivered the same way it always has been — no metered API, no subscription for the data itself. What changes here is who can find and evaluate it before that purchase happens. A coding agent scaffolding a wealthtech feature can pull a real sample and the schema without leaving its editor. An AI advisor agent — or whoever is evaluating one — can be scored against a fixed, cited answer key instead of a vibe. Neither requires a human to have read a landing page first.

Human, or otherwise — start here.

Browse the full catalog, or point an agent at the manifest and let it do the browsing.

Browse Wealth Data Sets Contact us