Skip to content

DFX Intelligence MCP

A lender asks who else is in the deal, at what spread, and when it matures. The answer exists, in a schedule of investments filed by each business development company every quarter, and assembling it by hand means reading the same borrower across a dozen filings and reconciling three spellings of its name. This server does that reconciliation and hands back the rows with the accession number attached.

Server type:  Remote (Streamable HTTP, JSON responses)
Endpoint:     https://exchange-production-9123.up.railway.app/mcp/core?c=dir-mcpso
Auth:         none for reads, optional account key for watch tools
Tools:        20 listed, 6 called live in this pass
Version:      0.16.0, serverInfo name dfx-real-estate, title DFX Intelligence
Protocol:     2025-06-18
Registry:     referenced in the submission, not re-probed here
Access tiers: anonymous screen depth, then a 7 day trial, then paid plans

What it does

The server is a read interface over a graph of private capital: private credit, real estate, family offices, independent sponsors, venture capital, private equity, registered investment advisers and public pension allocators. One tool, research, takes a whole question and plans the reads it needs, then answers from the rows it read. The other nineteen are those reads, exposed singly for drilling into any row.

Tool What it returns
research One question in one call: resolved entities, a direct answer built only from cited claims, coverage limits, what the tier withheld, paged claim rows
what_can_dfx_answer A plain objective in, the domain that answers it out, with the arguments and a free sample
resolve_name A firm, fund, person or company name to canonical ids across every graph, with the roles that matched
search_entities, get_entity Search by name or filter, then everything held on one id with relationships, dates and sources
search_relationships, relationship_path, get_capital_paths Who employs, manages, lends to, invests in or owns whom, and the evidence-backed path between two ids
changes_since What the graph learned since a time, driven by first sight rather than by filing date
verify A claim in, SUPPORTED, PARTIALLY_SUPPORTED, CONTRADICTED or UNKNOWN out, with the evidence rows
get_credit_provider, get_borrower_facilities A BDC or credit manager's quarterly tape, and every facility a borrower appears on
search_ria_advisor_moves, search_allocator_commitments Adviser registrations that moved firm, and public plans' own commitment disclosures line by line
get_research_records DFX's own published findings with permanent ids, versions and as-of dates
watch_create, watch_list, watch_changes, watch_reevaluate, watch_cancel A saved investigation with checkpoints and material-change deliveries, account key required

What this pass measured

A keyless initialize answered HTTP 200 with serverInfo dfx-real-estate 0.16.0, title DFX Intelligence, and tools/list returned 20 tools. Six were called and all six answered.

get_credit_provider on ARES CAPITAL CORP (CIK 1287750, file number 814-00663, New York) returned the tape as of 2026-09-30: 1,922 positions across 629 borrowers, fair value $29,212.6 million on the BDC's own SEC-filed total, principal held $23,063.0 million, cost $29,477.5 million, unfunded commitments $4,878.3 million, debt share 75.76 percent, first lien share 57.75 percent, 1,504 facilities, 54 sponsors financed, and 18 quarters on tape from 2021-12-31. The first quarter on that tape reads 1,112 positions over 390 borrowers, fair value $20,009.5 million, principal $16,384.4 million, mark on cost 0.9863, 121 PIK positions and an average spread of 6.27 percent. Every figure carries cite_url on sec.gov, the accession number and the source location inside the filing.

search_entities for Ares Capital resolved one record with claim_type VERIFIED_FACT and matched the adviser entity from 10-K text rather than from a name stem.

search_allocator_commitments with a query of CalPERS returned the plan's own disclosure rows: California Public Employees' Retirement System committed $2,000,000,000 to CV Consortio Fund L.P., managed by Valor Equity Partners, bucket private_equity, vehicle commingled, amount basis printed, cited to calpers.ca.gov.

research behaved differently on a resolvable question and an unresolvable one, and both are worth reading:

  • Named entity: the question about Ares Capital's fair value and first lien share ran three operations (resolve_name, get_credit_provider, search_private_credit_changes), returned state OK, and answered from the change tape rather than from a model: 2 new lender events, 2 new portfolio companies, 1 non-accrual placement, naming Flinn Scientific, Inc on the non-accrual and Essor Group Assets LLC and Aledade PBC as new portfolio companies.
  • No resolvable entity: the question about a dental services organization answered state UNKNOWN with the sentence "DFX could not resolve the subject of this question to one record, so it does not answer it", zero rows, zero claims, and llm_calls: 0. It declined rather than assembled a plausible paragraph.

Two more behaviours that matter before an agent is pointed at it. A search_allocator_commitments call with only a limit returned INVALID_ARGUMENT, charged: false, and the list of arguments it accepts, on the printed rule that it "refuses a call it cannot act on rather than guessing at what was meant". And verify checks a fixed predicate set (INVESTED_IN, ACQUIRED, CO_INVESTED_WITH, WORKS_AT, IS_A, ADVISER_OF, COMMITTED_TO, LENDER_TO, HOLDS_POSITION, MOVED_FIRM), so a sentence phrased outside that set comes back UNKNOWN with the reason, not a guess.

The withholding is stated, not implied

Anonymous access is real and it is limited, and the server says so inside every payload instead of failing at the door. A record names its subject and the first 3 related names per section; a list returns its first 5 rows in full and a count of the rest. On the Ares Capital record the locked block reported 29 related names withheld, and contact values and decision-maker names are never returned at any tier, only their types and counts. The server's own line is the one to quote to a user: LOCKED means withheld, not missing.

research runs at depth screen without a plan, and the response states that investigate, validate and monitor come with a paid plan or the 7 day trial. So the free path answers questions of the shape "what does the tape say about this named entity", and stops short of the deeper verification modes.

The caveats the server prints, which are the useful part

Each of these was read off the response, not inferred:

  • Principal on a facility or borrower is the sum of BDC pieces on the tape. It is a lower bound, never the commitment.
  • Fair value below cost is the lender's mark, not an impairment. Non-accrual comes only from a BDC's own footnote.
  • A maturity or spread that moved between quarter ends is an observed term change, not an amendment, until the amendment document is on the graph.
  • A sponsor is named only with its basis, and one-word borrower groups are never merged across BDCs.
  • A BDC that tags no maturity contributes nothing to the maturity wall, so the wall is a floor.

That list is why this entry earns a guide rather than a bookmark. A private credit number is only quotable when the reader knows whether it is a filing, a derivation or a lower bound, and this is the only server here that labels the difference in the payload.

Where this sits in the catalog

A grep of all 1,092 guides before this entry returned nothing for private credit, BDC, CMBS or allocator commitments, so the slice is empty rather than crowded. The neighbours are adjacent and stop short:

  • X1 Wealth MCP reads cited family office records, which is one investor type rather than the capital structure behind a borrower.
  • The institutional market data entries read public markets: prices, bonds, CDS and 13F holdings. A BDC schedule of investments is a private position with a fair value the lender itself assigns.
  • The EDGAR surfaces here read what a company reported about itself. This reads what its lenders reported about it, which is the opposite direction and the only way to see a borrower's whole facility stack.
  • Prometiam Risk and the company registry entries answer who a company is. This answers who is lending to it, on what lien, at what spread and until when.

Connect it

{
  "mcpServers": {
    "dfx-intelligence": {
      "url": "https://exchange-production-9123.up.railway.app/mcp/core?c=dir-mcpso"
    }
  }
}

The c=dir-mcpso parameter is the directory referral token in the vendor's submission. Expect it to change once the listing is processed, and take the canonical endpoint from the vendor's own agents page rather than from this guide.

Do not use it as

  • A credit opinion. The tape carries the lender's mark and the facility terms. It does not carry default, restructuring or non-accrual unless a BDC footnote said so, and the server says as much.
  • A commitment figure. Principal is a lower bound of what the BDCs on this tape hold, not what the borrower raised.
  • A name resolver across everything. Resolution is good, and the record states its basis; a name that matches one short word is not merged for a reason.
  • A free bulk export. Five rows and three related names per section is the anonymous ceiling, and the rest is a count.

FAQ

Does it need a key?

No for reads. initialize answered 200 with no credential and the six driven calls all answered. The watch tools need an account key, and their responses say so.

Can I quote an answer to a client?

Yes, with the citation that travels on the row: sec.gov URL, accession number and the section inside the filing. The claim basis field tells you whether you are quoting a filing, a derivation or a carried value.

Why did a question come back UNKNOWN?

Usually because it named no entity the graph could resolve, or because the claim used a predicate outside the ten the verifier checks. Both responses say which, and neither invents an answer.

Can it see my own portfolio?

No. The graphs are public filings, registrations and plan disclosures. The only writes are the caller's own saved investigations.

What does the paid tier add?

Full lists rather than five rows, related names past the first three, contact values where the vendor verified them, and the investigate, validate and monitor research depths. The trial is 7 days.

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

See Also

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