Skip to content

Tensorust Ontology MCP

Every due-diligence question is a join, and the models answer it without the join. A lender asks who controls the borrower. A procurement team asks whether a vendor appears on a banned list. A fund asks what other lenders mark the same loan at. A business owner asks whether the next acquisition is already over a regulatory threshold. Each of those is one answer standing on four or five public datasets that do not share an identifier, and an assistant asked cold will produce a plausible corporate structure that nobody recorded.

Tensorust is an ontology over those datasets with 2.4 million organizations, 1.16 million parcels, 428,254 FCC licences, 229,502 federal awards and 50,043 aircraft, joined on a canonical organization key, and it answers the join rather than the guess. What earns it a page is not the size. It is that every table carries an as_of date and a method paragraph naming the source file and the matching rule, and that it refuses a malformed argument instead of improvising, which the guides in this catalog do not get to assume.

First entry here for the ROAD Act investor-purchase screen, federal banned-vendor exposure over research papers, BDC private-credit mark disagreement, and the No Surprises Act arbitration market.

Server type:   remote Streamable HTTP, stateless
Endpoint:      https://shop.tensorust.com/api/mcp
Server:        tensorust-ontology, version 0.6.0
Protocol:      2025-06-18
Auth:          none for the free tier (no key, no account, no OAuth);
               a bearer key raises row limits
Tools:         41 across ownership, federal exposure, private credit,
               property, spectrum and arbitration
Repository:    https://github.com/Mattbusel/tensorust-skills
Docs:          https://shop.tensorust.com/mcp
Registry:      com.tensorust/ontology
Listed on:     the mcp.so submission queue as #7408

The evidence tier

Keyless and driven. initialize answered 200 as tensorust-ontology 0.6.0 on protocol 2025-06-18 with a full instructions field, tools/list returned 41 schemas, and 26 tools/call requests were run across 25 distinct tools, including four deliberate failure paths. No credential was used. Two of the tables below are dated by the server itself (as_of) rather than by me.

describe_schema is the first tool worth calling on any ontology, and this one returns its own inventory:

Object type Objects
Organization 2,446,357
Parcel 1,161,075
FCCLicense 428,254
Award 229,502
Aircraft 50,043
ControlGroup 48,569
ERateRequest 23,332
Sanction 19,958
PAC 3,331
Loan 3,214
Rule 7
FCCFiling 454

The ROAD Act screen, and the arithmetic that makes it a product

road_monitor_summary returns the rule before it returns the data: the 21st Century ROAD to Housing Act section 1001, at 42 U.S.C. 19521, enacted 2026-07-11, effective 2027-01-07, proposed rule 31 CFR Part 36 under RIN 1505-AC96, docket TREAS-DO-2026-0529, at 91 FR 66799, comments due 2026-11-23, threshold 350 homes, penalty up to the greater of $1,000,000 or 3x the purchase price per purchase.

Then the population: 43 covered control groups holding 166,797 homes, of which 28 sit in the 350 to 700 band, 19 hold more than 350 homes with no single entity above 350, and 6 are hidden from a name search entirely (3,442 homes held by groups a name query does not surface), across 2,325 covered legal entities.

road_investor_group on Invitation Homes returns the shape of the problem in one object: 32,315 homes in coverage, 251 legal entities, sixteen counties from Bexar to Travis, a largest single entity of 2,079 homes (SRP SUB LLC), and over_threshold_by 31,965. The check is one subtraction: 32,315 minus the 350 threshold is 31,965. A 251-entity structure where no entity crosses the threshold is exactly why a group-level screen and an entity-level screen disagree, and the link_kinds block shows where the structure was recovered from: 180 root links, 90 parent, 81 mailing-address and 14 SEC.

road_purchase_stream extends it to transactions: 19,688 purchases in the stream, the newest from Mecklenburg County NC on 2026-11-02 with county, parcel id, situs address, buyer entity, the buyer's own home count, the group, the group status and a signal string, so a single purchase can be read as BUYER_ENTITY_UNDER_350 inside a COVERED group. That is the exact case the 350 test was written around, returned as a field.

One inconsistency recorded rather than smoothed: road_monitor_summary reports 43 covered groups while rule_coverage on ROAD-1001 returns a count of 64 items and lists 64 of them. Two different definitions in two tools on one server, and a caller who quotes one without naming which will be wrong by twenty-one groups.

Private credit: what two lenders say the same loan is worth

credit_disagreements is the strongest table in the whole surface. It is dated 2026-06-30, covers 2,189 companies held by 2 or more managers, and returns 71 where the marks differ by ten points or more, with the source named as BDC schedules of investments, first-lien term loans, fully funded, cost 85 to 110 percent of par.

The first row: Assembly Intermediate LLC, an 80.2 point gap, lowest mark 0.198 (Goldman) against highest 1.000 (Morgan Stanley) on $53,825,000 of par across two funds. The second: Norvax LLC, 69 points, 0.1651 against 0.855 across three funds, par $4,366,000. Check the gap: 0.855 minus 0.1651 is 0.6899, which is 69 points to the nearest whole. lender_marks on the same company returns the fund-level rows underneath (Blue Owl Capital Corp at 0.1649 and 0.1651, Midcap Financial at 0.855), the 2029-08-05 maturity, and a backtest sentence with its own numbers: lenders sitting ten to twenty points above the lowest other manager cut their own mark by five or more points the next quarter 20.4% of the time, against 2.8% when they agreed.

credit_q3_cuts shows the same table moving: dated 2026-09-30, 123 funds filed, last filing 2026-11-13, 209 cuts of five points or more. The largest is Wash And Wax Holdings LLC, from 0.9514 to 0.1684, which is -78.3 points, and the arithmetic is the subtraction. contractor_credit_watch then joins marks to federal money: of 202 contractors, 28 are stressed and carry $6,748,062,967 of federal FY2025 obligations, with a private-only cut of the same population at 150 contractors and 19 stressed.

Federal exposure: the banned-vendor question, with the method attached

exposure_summary is dated 2026-11-13T11:32:16Z and opens with the method rather than the answer: a paper counts only when Europe PMC finds the vendor in its methods, results, figures, tables, supplement or acknowledgements and a sentence names it; engineering and physical-science papers come from the OpenAlex full text; author affiliations and reference lists never count; grant numbers are matched exactly to USAspending FAINs. The tiers are named (broad, and the narrower one), which is the difference between a screen and a rumour.

perf_summary is dated 2026-11-06T03:54:03 and names its own inputs by file and date: USAspending award archive FY2025 and FY2026 with FY2027 to date, DOL EFAST2 Form 5500 and 5500-SF bulk filings received 2024 to 2025, FSRS subawards, and 13 CFR 121.201 size standards from the eCFR API on 2026-10-30. Its capacity method divides set-aside obligations by the prime's own active participants on its largest Form 5500 plan, with the exclusion list spelled out (D?1x as-a-Service, V2xx travel, construction and specialty-trade NAICS).

who_owns_home answers the property question or says it cannot: 1 Infinite Loop, 95014 came back covered: false with the note Outside the counties covered so far, which is the honest version of a 17-county coverage limit rather than a zero that reads as no owner.

The refusals, which is why this one is quotable

Four failure paths were driven and each answered as a failure rather than as a guess:

  • idr_offer with code: 11 refused: code must be a CPT or HCPCS code, e.g. 99284, 99285, 36415, A0427. The code space was validated, not coerced.
  • erate_deadline_check on BEN 16012345 returned No funded E-Rate requests from funding year 2023 on, and no Category 2 budget, rather than a deadline for a nonexistent applicant.
  • perf_firm on Booz Allen Hamilton returned found: false with Not a set-aside recipient with $250K+ in FY2025-27 to date, which is a scope statement and not a claim that the company does not exist.
  • get_dataset_sample on private-credit-marks returned isError: true and has no public sample file yet instead of an empty sample that would read as a dataset with no rows.

Where the surface is thin, measured rather than assumed

aircraft_owner with owner: APPLE returned 13 rows and every one was a substring match: APPLEBEE AVIATION INC, APPLE REALTY GROUP INC, and other names that merely contain the four letters. The FAA lookup is a name search, not an entity resolution, so pin it with an N-number when the question is does Apple own an aircraft. exposure_awards with q: BGI returned 0 rows across every counter, which is a naming problem rather than an absence, and the same tool accepts a vendor argument for exactly that reason.

idr_scorecard is a market table rather than a lookup: CMS Federal IDR Public Use File 2025 Q4, altered and aggregated by Tensorust and labelled as such. FHAS carries 135,347 disputes at a 525 dollar single fee, a 77.5% provider win rate on disputes, a median prevailing multiple of 3.72, a 20 day median and $61,454,979 of fees in the quarter; PRI carries 60,796 disputes at 518 dollars, a 94.3% win rate, 4.43 and 47 days. The fee totals are the vendor's arithmetic on a two-sided fee model and are not reproducible from the dispute count alone, so they are quoted as the server's figure.

Limits

The property layer covers 17 large counties and says so at the point of failure rather than in a footnote. The ontology is US-only. The free tier truncates to samples and row limits, so a full export is a paid dataset product, and the offers sit on the same host. Coverage dates differ per table (June 30 for marks, September 30 for cuts, November 6 and November 13 for the exposure and performance builds), so two answers from two tools are two vintages and should be labelled that way.

FAQ

Does it need a key?

No. initialize, tools/list and the read tools answered without one, with sample rows and row limits. A bearer key raises the limits, and the dataset catalogue on the same host sells the full extracts.

What makes it different from a search engine?

The join and the date. A search returns a page about a company; this returns the control group, the federal awards, the FCC licences, the private-credit marks and the rules that apply, from one key, with the build date of each table printed in the answer.

Can I rely on the ROAD Act numbers?

The statute, effective date, threshold and penalty are quoted with citations you can check, and over_threshold_by reconciles by subtraction against the stated threshold. The one number to watch is the covered-group count, because road_monitor_summary says 43 and rule_coverage says 64, so name the tool you are quoting.

Where does this fit a business?

Procurement screening, credit and valuation checks, real-estate acquisition diligence, and the regulatory answer to is this investor already over the line are the four places this beats a chat answer. It is the first entry in this catalog for all four.

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