The Lunenburg, Massachusetts town and school budget, every figure traced to its source document.
Scored across 8 tools
Each tool targets a distinct concern: dataset discovery, data-grain warnings, safe budget lookups, staff counts, meeting full-text search, provenance, SQL catch-all, and query templates. The one soft boundary is that `query` could theoretically reproduce `budget_history` and `staff`, but the descriptions aggressively warn against naive SQL for those cases, which makes the specialization meaningful.
`list_datasets` and `search_meetings` follow a verb_noun pattern, but the rest mix bare nouns (`document`, `query`, `staff`), compound nouns (`budget_history`), an imperative phrase (`read_first`), and an adjective_noun (`worked_examples`). All names are readable and memorable, but no consistent syntactic convention holds across the set.
Eight tools is well within the ideal 3-15 range for a municipal archive server. Each tool earns its place: discovery, pre-query guidance, arbitrary SQL, two specialized wrappers that exist specifically to prevent costly analytical errors, full-text search over a separate corpus, provenance, and a curated set of executable examples.
The domain is well covered: dataset discovery, data pitfalls, arbitrary read-only querying, meeting document search, and citation. The catch-all `query` tool avoids most dead ends. Minor gaps remain — no full-text search over non-meeting documents like town reports and no bulk export/download dataset tool — but the 250k-row guardrail and publisher URLs partially mitigate these.
What a school budget line was in each year, AT ONE STAGE. The stage argument is required and singular on purpose: proposed, settled and actual are three different documents about the same year, and a growth rate measured from an actual to a budget is partly growth and partly the step between them. That mistake put a special education escalator 1.5 points too high here and was invisible until somebody asked how the number was derived. Note also that a budget line is NET — what the town must raise after grants, fees and state aid — so a line can rise because a grant ended rather than because anything cost more.
Inputs: label (required), stage (required).
Where a document came from: the publisher's URL, our copy, and its sha256 so a reader can check they have the same bytes. Use it to cite anything.
Inputs: name (required).
Every dataset in the archive, with THE YEARS EACH COVERS, its row count and its size. Call this before answering from prose: it is how you find out whether the archive holds data for the year and subject you are being asked about. 49 datasets covering the town and school budgets, the town ledger, staff rosters, out-of-district placements, elections, and fifteen years of annual town reports.
Inputs: no named parameters.
Any question the other tools do not cover, as one read-only SQL statement over the archive database. SELECT or WITH only; a LIMIT is imposed if you omit one. Call read_first before computing anything and list_datasets to find table names. A query estimated to read more than 250,000 rows is refused — narrow it with a WHERE, or ask for one table at a time.
Inputs: sql (required).
The grain of every table and the specific ways to get a confident wrong answer out of this data. Read this before computing anything. It states, among others, that a budget and an actual must never be combined in one calculation; that a budget line is NET of grants and fees and is not what a thing costs; and that no budget line is mapped to a ledger account, so budget-to-actual at line level cannot be answered from this data at all.
Inputs: no named parameters.
Which of the 1,422 published town meeting documents contain a word — every board, 2025 onward. Returns the board, the date and a citable URL for each. AN EMPTY RESULT MEANS THE WORD IS NOT IN THE INDEXED DOCUMENTS, which is not the same as nobody having said it: the archive starts in January 2025. It matches words exactly, so plurals are separate terms — search "jersey" and "jerseys" both.
Inputs: word (required), board.
How many people the town PRINTED on a school staff roster, by year, school and kind of job. Uses our classification of the printed title, never the title itself, because the town has called the same job Tutor, Aide, Paraprofessional, Para, (para) and Sped Para across fifteen years. THIS IS A COUNT OF NAMES, NOT A STAFFING LEVEL: a roster carries no FTE, so a 0.4 music teacher and a full-timer are one row each, and it names no funding source, which is the question that usually matters. Grade appears only where the page happened to print it.
Inputs: fy, category.
107 questions this archive can answer, each with the SQL that answers it. Every one is executed against the database on every build, so none of them is a claim. Start from the nearest one and edit it rather than writing a query from scratch.
Inputs: no named parameters.
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
The public record of local government: what your officials actually say and do, word for word.
Financial data, fiscal health scores, and economic indicators for 20,000+ U.S. municipalities.
SACS California school finance: public district financial data, CDE filings, comparisons, briefings.
Track budgets from your AI chat: envelopes, monthly limits, variance reports. All data is local.
Add one secure layer between your agents and this server.
