Skip to content

NYSgpt Data Library

A question about federal spending, patents or scholarly output is a question about a table nobody's model has read. The records exist, they are public, and they are far too large to be in a prompt: USAspending, PatentsView, OpenAlex, SEC filings, LEHD jobs, EPA emissions, FERC sales, Medicare payments. The useful shape for an assistant is not a document to read. It is a SQL endpoint with the tables already loaded and a schema it can inspect before it writes a statement.

NYSgpt Data Library is that endpoint, and it is keyless: list_databases, list_tables, describe_table and a read-only query in the Athena (Trino) dialect, free without a key up to a daily limit. This pass called all four tools through nine requests and measured what a caller can actually rely on: 36 databases, a patents database of 128 tables holding 3,055,846,513 rows, a query that returned in 1,400 bytes scanned, a clean refusal of a write statement, and two defects worth knowing before an agent is pointed at it.

Server type:   remote Streamable HTTP
Endpoint:      https://data.nysgpt.com/mcp
Server:        nysgpt-data-library, version 2.0
Protocol:      2025-06-18
Auth:          none for the free tier (30 queries a day per the submission; paid keys as X-Api-Key)
Tools:         4 (list_databases, list_tables, describe_table, query)
Website:       https://nysgpt.com/data
Listed on:     the mcp.so submission queue as #7487
Registry:      com.nysgpt/data-library 2.0.0

The evidence tier

Keyless and driven. initialize answered 200 as nysgpt-data-library 2.0 on protocol 2025-06-18 with no credential. tools/list returned four closed schemas, and all four were called: list_databases once, list_tables twice, describe_table twice, query four times (two of them deliberate failures). The free-tier allowance was not exhausted.

The tool surface

Tool Arguments What it returns
list_databases none The databases, one per subject, each with a one-word description.
list_tables database Every table in one database with its row count and table type.
describe_table database, table Field names, types and descriptions, the table description and the row count.
query sql One read-only statement, up to 1,000 rows, with column names, bytes_scanned and a query_id.

The catalogue, measured

list_databases returned 36 databases: awards, census, consumer, courts, economy, education, elections, energy, environment, executive, fec, finance, geo, grants, health, housing, insurance filings, labor, law, linked, nonprofits, oversight, patents, reactors, reference, regulations, research, scholarly, sec, siting, solar, tax, trade, trademarks, treasury, vehicles.

list_tables on govblock_patents returned 128 tables whose stated row counts sum to 3,055,846,513. The largest are current enough to be worth naming:

Table Rows
patex_transactions_25 504,754,561
patex_attorney_agent_25 366,993,004
pv_g_us_patent_citation 152,631,929
claims_pgpub_fulltext 82,017,562
cpc_master_pgpub 71,736,962
cpc_master_grant 70,324,580

The _25 suffix on the prosecution tables is the year: these are PatentsView prosecution datasets, not a single concatenated archive, which is the difference between a schema an agent can bound and one it has to scan.

One table, read end to end

describe_table on govblock_patents.hist_annual reported 175 rows and 147 fields, every one typed string, with an empty description on all 147. Read literally that is a weak schema: a caller has to know that total_app means total applications and that nber11_inforce is a NBER class count in force. The table description is not prose either; it is the source path the loader recorded, govblock-lake-638175140432/lake/v1/patents/raw/histexc//annual.csv$.

The query is where the table earns its use:

SELECT year, total_app, total_inforce
FROM govblock_patents.hist_annual
ORDER BY year DESC
LIMIT 4
year total_app total_inforce
2014 412,861 2,565,874
2013 402,538 2,411,756
2012 381,508 2,260,800
2011 357,132 2,130,053

The row count and the coverage agree. The newest year in the table is 2014 and the table holds 175 rows, which is exactly the span from 1840 to 2014 inclusive. So this is the historical series, not the current one, and a caller asked for the newest patent count must go to patex_* or the grants tables rather than this one. The response reported bytes_scanned: 1400 and a query_id, which is the evidence that the statement executed against Athena rather than being served from a cache of somebody's answer.

The refusal, and the defect beside it

A write statement is refused cleanly. DELETE FROM govblock_patents.hist_annual WHERE year='2024' returned Error: only SELECT, WITH, SHOW and DESCRIBE are allowed. The refusal arrives as HTTP 200 with the sentence in the tool result rather than as a JSON-RPC error, so a client that only inspects the transport status will read a refusal as success.

An unqualified table name surfaces an AWS internal error and an account identifier. The same query with hist_annual instead of govblock_patents.hist_annual returned query FAILED: Insufficient permissions to execute the query. User: arn:aws:sts::638175140432:assumed-role/query-api-lambda/nysgpt-query-api is not authorized to perform: glue:GetDatabase on resource: arn:aws:glue:us-east-1:638175140432:database/default. The tool's own description says to qualify every table, so the rule is documented; what is not documented is that breaking it fails on a permissions path in the default database and hands the caller the AWS account number and the execution role's name. It is a defect to record, not a security incident: the identifier belongs to the vendor's own account and grants nothing.

Two surfaces of one vendor that do not agree on the size

The registry record com.nysgpt/data-library 2.0.0 describes the corpus as 52.8B public records. The server's own initialize instructions say 69 billion. The submission to the queue says 52.8 billion records in 6,396 tables. Three surfaces, two numbers, one vendor, and nothing in the tool output reconciles them, because the only size a caller can measure is the per-database table counts. This pass measured one database: 3.06 billion rows in patents alone. A cumulative total across 36 databases is not something a caller should quote from a marketing line, and this guide does not.

Is this worth a connection?

For an operator asking real questions about public money, patents or research output, yes, and the reason is the schema pass: list_tables plus describe_table is the sequence that lets an agent write a bounded statement instead of guessing table names, and bytes_scanned on every answer is the feedback a cost-conscious caller needs. The free tier carries no credential at all, which makes it usable in a cron or a shared agent without a secret.

What it is not: a curated dataset. Fields are text, descriptions are empty in the tables sampled, and the coverage of the historical series ends in 2014, so the useful pattern is describe first, query second, and never assume the newest year is present. The paid tier, the exact daily allowance and the paid-key header are vendor statements that this pass did not test.

FAQ

Does the NYSgpt Data Library need an API key?

Not for the free tier. initialize, tools/list and all nine tools/call requests ran with no credential, and the paid path is an X-Api-Key header or a ?key= parameter.

What SQL dialect does it speak?

Amazon Athena, which is Trino. The tool description warns that many source columns are stored as text and need try_cast, and the patents table this pass read typed all 147 of its fields as string, so cast before aggregating.

How do I find the right table?

list_databases, then list_tables on the database, which returns each table with its row count, then describe_table before writing the statement. Qualify every table as database.table; an unqualified name fails on the default database.

Why did a query return 2014 as the newest year?

Because that table is the historical series. The table holds 175 rows covering 1840 to 2014, and current-year counts live in the patex_* and grants tables instead.

Can it write, update or delete?

No. Only SELECT, WITH, SHOW and DESCRIBE are accepted, and a DELETE returns that sentence as the tool result.

Integration with CorpusIQ

CorpusIQ is the connector layer that lets AI assistants read the systems a business already runs on - accounting, payments, CRM, email, files and analytics - and answer with the source named. It is read-only by design: it retrieves data and never writes to, or acts in, a connected system, which is what makes it safe to point at live accounts.

This page is part of the CorpusIQ directory, the reference we maintain for the MCP server and agent-skill ecosystem. Run alongside CorpusIQ, a third-party server like this one can be combined with a company's own business data in a single answer.

See what CorpusIQ connects

AI answers on your data. See where every number comes from.

Connect your tools and ask in ChatGPT, Claude, or Perplexity. Enterprise-grade answers in minutes. From $29.95/month.

Start free, no credit card
Contact Terms Privacy Pricing