Release notes

The site is built directly on the U.S. Department of Labor EBSA public Form 5500 Datasets. Ingestion is automated and doctrine-bound: no hard-coded column names — every field is resolved through an alias table and anything unmapped is preserved in a raw jsonb column — a streaming parser for the large yearly files, and resumable, idempotent runs. Every run is recorded in ingest_runs and every surprise in ingest_warnings. Each filing is keyed by ACK_ID, the join key across every schedule.

• plan_filings — the filing record, one row per filed Form 5500 / 5500-SF.
• sch_c_providers — Schedule C service-provider compensation, direct and indirect.
• sch_h_lines — Schedule H financial line items.
• sch_h_compliance — Schedule H Part IV self-reported compliance answers and the accountant opinion.
• dataset_meta, ingest_runs, ingest_warnings — dataset vintage tracking and the audit trail.

The server exposes nine public tools with no login or key required:

• search_plans — find plans by sponsor, EIN, state, or size.
• get_plan — the full filed profile for one plan.
• top_plans — rank plans by a filed metric.
• plan_changes — filed-value differences across a plan's filing years, separating a real change from a first capture or a restatement.
• cohort_stats — percentile distributions and base rates for a defined cohort in one call.
• cohort_chart — the same distribution as a percentile box-plot image with one plan marked.
• list_skills and get_skill — discover the guided workflows this server ships and load one's full instructions at the moment it is needed.
• server_info — the server's live version, tool list, and skills-catalog state, so an assistant can detect a stale client cache itself.

New capability ships as a new skill or an optional parameter — tool names and their input meanings do not change, so a connected client never breaks.

Skills are written procedures for Form 5500 work — what to pull, in what order, and what not to say. A connected assistant reads them from the server via list_skills and get_skill; the catalog currently stands at 9:

• Form 5500 literacy — What each filed Form 5500 field means, how large-plan and small-plan filings differ, and which conclusions the data does and does not support.
• Presentation contract — The house rules for rendering any Form 5500 output — lead with the answer, tabulate parallel figures, keep fee positions and caveats honest, and collect caveats at the end. Layout only; it never changes a filed figure.
• Plan fee benchmark — Place one plan's filed service-provider compensation against a defined peer cohort as percentiles, with the cohort definition, cohort n, and filed-fee coverage stated.
• Plan diligence brief — A one-page read on a single plan built only from its filing: size, type, funding, service providers, filing-history changes, and the questions the filing cannot answer.
• Fee transparency report — Itemize every service-provider fee a single plan filed - Schedule C provider-by-provider compensation and Schedule H administrative-expense lines - laid out side by side and reconciled, exactly as filed, with no benchmark and no judgement.
• Compliance flag scan — Read a single plan's Schedule H Part IV self-reported compliance answers and the independent-accountant opinion exactly as filed, present each disclosed item as a fact and the question it raises for the sponsor - never as a finding or accusation - with optional cohort base rates.
• Committee review inputs — Assemble the filed inputs a retirement-plan committee needs to review one plan - itemized fees with cohort position, funding and financials, the Schedule H Part IV compliance disclosures and accountant opinion, and the questions to put to providers and counsel. Inputs to a reasonableness review, never a conclusion.
• Verify a plan claim — Check a public claim about a plan - a recordkeeper's pitch, a sponsor's statement, a headline - against the plan's filed Form 5500 record, returning one of four verdicts and naming any definitional gap before reporting one. Never calls a claim false or misleading.
• Plan prospecting screen — Screen the filed Form 5500 population for plans matching typed criteria - size, plan type, fee presence, disclosed Part IV items - and return a neutral shortlist with cohort context and the questions each filing raises. A discovery tool over public filings, never a judgement about any plan.

Every skill carries explicit guardrails. They report figures exactly as filed, never estimate a number a plan did not file, position fees as percentiles within a stated cohort rather than calling them high or low, and treat a self-reported compliance answer as a disclosure and a question, never an accusation.

The presentation-contract skill is a house rendering standard every guided workflow now loads. Three rules: lead with the answer; put any set of three or more parallel figures — Schedule C provider compensation, Schedule H expense lines, cohort percentiles, compliance base rates — in a table rather than a sentence; and collect the caveats into one closing block. It carries the discipline this data already holds — a field a plan did not file reads "not filed on this form", a fee position always travels with its cohort n and filed-fee coverage n and is never called high or low, derived ratios are labelled, and the Schedule C-vs-Schedule H gap is named as expected rather than an error. Layout only; it never changes a filed figure.

It is wired into all seven workflow skills. Skills are served fresh from the database on every call, so this reached every connected client with no reconnect.

Every skill now carries a machine-readable output_contract — a format plus the ordered sections its answer should follow — returned by both list_skills and get_skill, matching the advMCP.ai sibling. The stored contract moved from free text to structured JSON, and the skills pages render it.

Schedule H covers roughly 87,500 large-plan filings with about a million filed line items — total assets, contributions, distributions, and the administrative-expense lines (investment-management, contract-administrator, other-administrative, and total administrative expenses). A plan without a Schedule H shows "Not filed" for assets rather than a substitute figure.

get_plan returns a compliance block for every large-plan filing: the Part IV self-reported answers — delinquent participant contributions (line 4a), non-exempt party-in-interest transactions (4d), loans and leases in default (4b/4c), loss from fraud or dishonesty (4f), fidelity-bond coverage (4e), and assets with no readily determinable value (4g) — plus the independent qualified public accountant (IQPA) opinion (unqualified, qualified, disclaimer, or adverse), the audit scope, and the accounting firm. Each answer is the plan's own, exactly as filed; a blank is "not filed", never No and never zero.

A self-reported "Yes" is a disclosure the plan itself made and a question for the sponsor — never a finding, a violation, or an accusation, and the guided compliance-flag-scan skill is built to never use those words.

cohort_stats includes a compliance_base_rates block: for a defined cohort, the share of plans that answered each question Yes, each with its own answered-n and a coverage count. Across the large-plan population, about 20% disclose a delinquent-contribution item and about 3% carry a non-unqualified audit opinion.

On line 4a the DOL dataset carries only the Yes/No — there is no delinquent dollar amount in the file — so the amount is reported as "not filed" rather than a substitute figure.

Coverage: 99,846 Schedule H filings across the 2024 and 2025 vintages. Schedule H — and therefore Part IV — is a large-plan schedule; 5500-SF filers do not file it.

A resumable pipeline refreshes the DOL data monthly via pg_cron — the DOL republishes recent filing years around the first of each month. The Schedule H financial and compliance feeds refresh on the same schedule, an ingest pass then a derivation pass, both idempotent, so a vintage already loaded is a no-op and an amended filing is re-typed. If the current vintage is already loaded, the run no-ops.

Row-level security makes the filed data publicly readable while keeping ingestion and rate-limiting to the server-side service role. The MCP endpoint is intentionally open for read-only access to public DOL records.

The brand is 5500mcp (5500mcp.ai), named after the form it serves — a sibling to advMCP.ai for SEC Form ADV. Dark mode by default, primary navigation in the footer, theme-aware logos, and legal pages, a sitemap, and robots.txt for discoverability.