DFX Intelligence MCP server

# read the schemas, no handshake required
curl -s https://exchange-production-9123.up.railway.app/mcp

# call a tool
curl -s https://exchange-production-9123.up.railway.app/mcp \
  -H 'Content-Type: application/json' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":
       {"name":"resolve_address","arguments":
        {"address":"100 Binney St","city":"Cambridge","state":"MA"}}}'

Registry entry: io.github.Capital-W-Holdings/us-property-parcel-real-estate-debt. Transport is Streamable HTTP. Both the current 2026-07-28 revision and the older initialize handshake are served, because most deployed clients still send the latter.

Claude Code. One command, no config file. Verified against this endpoint: the server reports Connected on the next claude mcp list.

claude mcp add --transport http dfx-real-estate https://exchange-production-9123.up.railway.app/mcp

Claude Desktop. Settings, then Connectors, then Add custom connector, and paste the endpoint. On a build with no custom connector option, put this in claude_desktop_config.json instead and restart the app. The bridge exists because that config accepts a subprocess rather than a URL; it is not a DFX requirement.

{
  "mcpServers": {
    "dfx-real-estate": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://exchange-production-9123.up.railway.app/mcp"]
    }
  }
}

Cursor, Windsurf and Cline. A remote server is a URL entry in ~/.cursor/mcp.json, .cursor/mcp.json for one project, or the equivalent file in the other two.

{
  "mcpServers": {
    "dfx-real-estate": { "url": "https://exchange-production-9123.up.railway.app/mcp" }
  }
}

VS Code. .vscode/mcp.json in the workspace, which uses servers rather than mcpServers and wants the transport named.

{
  "servers": {
    "dfx-real-estate": { "type": "http", "url": "https://exchange-production-9123.up.railway.app/mcp" }
  }
}

Anything else that speaks MCP over stdio only. npx -y mcp-remote https://exchange-production-9123.up.railway.app/mcp puts a local stdio process in front of the HTTPS endpoint. No key is passed through it because there is no key.

There is no account to create and no credential to rotate, so an install that works once keeps working. If a client asks for an API key before it will save the server, the field can be left empty or filled with any string: the endpoint does not require one and never checks it.

Read from tools/list when this page was rendered, so the table cannot say a number production does not. Real estate first, then the family office, independent sponsor and venture capital tools, then the tools that work across every domain.

Family Office Intelligence. 2,569 offices on the graph (695 candidates, 157 confirmed multi-family, 336 probable single-family), 4,238 foundations, 748 offices with 13F positions, 410 with observed direct investments.

Independent Sponsor Intelligence. 4,189 sponsors, 10,726 capital providers, 97,287 private companies with plan-filing history of which 17,079 carry a transition signal, 317,710 computed company-to-sponsor matches with reasons and blockers, 2,407 announced transactions.

Venture Capital Intelligence. unknown firms (unknown with a fund raising in the last 18 months), 77,255 funds with every fund amount kept apart, 348,556 people, 144,324 companies, 145,614 rounds. Stated sectors are populated on unknown firms today, so a sector filter on firms answers NOT_COVERED rather than an empty list.

Across all of them: search_entities, get_entity (everything on one id: card, relationships, events, evidence, cross-graph links), search_people, search_relationships, relationship_path (how X connects to Y, every hop an evidenced edge), search_events, verify (SUPPORTED, PARTIALLY_SUPPORTED, CONTRADICTED or UNKNOWN, with the observations), and the economic tools find_capital_for_opportunity, find_opportunities_for_capital, explain_match (reasons and blockers, never a bare score), why_now and who_should_care.

Contact points are withheld over MCP on every graph. Call what_can_dfx_answer with domain set to family_office, independent_sponsor, venture_capital or real_estate for that domain's entity types, event families, rights, freshness and limitations.

Resolving 100 Binney St, Cambridge MA returns a parcel with its assessed value and a dfx_id. Passing that id to get_property_record returns the recorded sale:

{
  "object": { "type": "parcel", "name": "100 BINNEY ST",
              "city": "Cambridge", "state": "MA",
              "assessed_total": 494419400.0, "assessment_year": 2026,
              "address_ambiguous": false },
  "sales": [{
    "transaction_type": "sale",
    "closed_date": "2022-04-01",
    "consideration": 1020000000.0,
    "allocated_consideration": 1020000000.0,
    "allocation_basis": "single parcel deed, consideration as published",
    "parcels_on_this_instrument": 1,
    "arms_length": "unknown",
    "arms_length_basis": "$1,020,000,000 is 2.06x the FY2026 assessed value…",
    "recording": { "book": "79922", "page": "0115" }
  }],
  "coverage_notes": ["Recorded sales here carry a buyer but never a seller…"]
}

Both consideration figures travel together deliberately. A deed repeats its full price on every parcel it covers, so summing the naive column multiplies a portfolio sale by the number of parcels in it.

# 1. establish a position. returns no events, on purpose.
{"name":"changes_since","arguments":{}}
  -> { "count": 0, "first_call": true,
       "cursor": "2026-08-31T10:38:55.382135+00:00" }

# 2. later, ask what arrived. pass the cursor back unchanged.
{"name":"changes_since","arguments":{"since":"2026-08-31T10:38:55.382135+00:00"}}
  -> { "count": 50, "complete": false,
       "cursor": "2026-08-28T23:23:06.51893+00:00|30adacd5-36e3-4cd8-8752-0e4377353105" }

# 3. keep calling with the newest cursor until complete: true.

The argument is since and the response field is cursor. They are the same value and the names do not match. A client that reads cursor and sends it back as cursor gets a first-call response with zero events and no error, which looks like “nothing changed” forever. It is the first thing a careful caller gets wrong, so it is printed here rather than left to be discovered.

Paging is exact. The cursor is a composite of the learn-time and the event id, so a page boundary falling inside a group of events learned in the same instant does not skip the remainder. Until 2026-08-31 the cursor was the timestamp alone, and because the tape carries 58,743 events across only 254 distinct learn-times, an agent paging to completion saw roughly a fifth of it and was told the tape was complete. Fixed, and pinned by a test that walks tapes with tie groups of 1 to 500 rows at page sizes 1, 7 and 50 and asserts every row is returned exactly once.

An empty response returns your own cursor unchanged, because standing still is safe where advancing is not. A malformed cursor is refused as INVALID_CURSOR with safe_to_retry: false, never as a DFX outage.

An empty list is a claim, and this endpoint will not make it carelessly. Five distinct outcomes are served rather than collapsed:

• Known no result. DFX looked inside applicable coverage and found none.
• Coverage unavailable. The geography or category is not held. Says so.
• Resolution failed. The input could not be keyed; the reason is returned.
• Rights blocked. DFX holds the information and may not republish it.
• Temporarily unavailable. A DFX availability problem, marked safe to retry, explicitly not a statement about your query or our coverage.

Every observation carries its source, the time it was retrieved and a content hash. Rights are determined by reading the operative terms for each source and are recorded as versioned rows: when a determination changes, the old one is superseded rather than edited, because the question diligence asks is what the terms said on the day we ingested.

Unclear means no. Two sources totalling over two million observations are held and not published because their terms could not be established. Sources published under an attribution condition carry that notice in the response envelope itself rather than in documentation somebody may not read.

Measured cold path from a standing start with no prior knowledge: five requests, about 1.5 seconds total, to a first useful answer. Typical single tool call is 200ms to 700ms.

All 134 tools are free, read-only and unauthenticated. No key, no signup, no OAuth, and no account or payment step anywhere on the server.

The loan tape is free too: debt_maturity_schedule returns up to 200 loans per call for one state. It is the loan rather than the event: the principal, the lender, the instrument and the origination, rather than a date and a headline. It also reaches loans the event index does not, which is the gap described above. Every response states how many loans DFX holds for the filter, so a capped answer is never presented as the whole population.

How many rows a state actually holds, because it is not the same everywhere. 19,881 loans carry a maturity date and every one of them carries maturity_basis: confirmed, filed with the SEC by a servicer or recorded by HUD. Not one is estimated or inferred from a term length. But only 1,614 of them fall inside the default eighteen-month window, and they concentrate: California holds 329, New York 193, Texas 117, Ohio 61, and twenty-eight states hold fewer than twenty each. Montana and Wyoming hold one apiece. Widen within_days to reach further out.

Bulk access and higher volumes. If you need the full tape, higher limits or a feed, write to jesse@dfxintel.com and say what you are building.

Rate limits, so you can plan around them instead of discovering them. 20 requests per minute and 300 per hour for any caller. search_property_events and changes_since are additionally capped at 60 per hour, because those two are the paging capabilities. A refusal is a structured answer carrying the state, the reason and a retry time rather than a bare 429. Search results cap at 50 rows per call; the loan tape caps at 200.

Each of these answers one question, states the coverage for that question alone, and shows a real call with the real response. They exist because an agent searching for a capability should land on the capability rather than on a reference manual.

• Which commercial real-estate loans mature in a given state and window?3,422 published events, 52 states and territories, effectively national. Free.
• Which LIHTC properties are reaching the end of a compliance period?11,956 published events, 56 states and territories, effectively national. Free.
• Which HUD-subsidised properties have contracts approaching expiry?4,721 published events, 54 states and territories, effectively national. Free.
• Where is commercial real estate in distress, foreclosure or workout?307 published events, 28 states, broad but NOT national. Free.
• What did this property sell for, and who owns it?43,680 published events, 2 states only: MA, NY. Free.
• Which commercial leases are approaching expiry, and who occupies a building?3,966 published events, 55 states and territories, effectively national. Free.

Every count above was measured against the running exchange on 12 September 2026. These pages are built rather than fetched, so a count is a dated reading and not a claim about this moment; call dfx_coverage for the grid as it stands when you read it.

Everything on this page is available without reading it. None of these need a key or a handshake:

• GET https://exchange-production-9123.up.railway.app/mcp returns the full tool schemas, the coverage block and a worked example.
• GET https://exchange-production-9123.up.railway.app/openapi.json is the OpenAPI description of the same surface.
• GET https://exchange-production-9123.up.railway.app/.well-known/agent-card.json is the A2A agent card.
• GET https://exchange-production-9123.up.railway.app/.well-known/mcp/server-cards.json is the SEP-2127 server card.
• GET https://exchange-production-9123.up.railway.app/ returns the capability manifest with versions and coverage.

GET https://dfxintel.com/ai/real-estate-mcp/catalog.json is the full machine-readable capability catalog: every tool schema, the measured per-family coverage grid, the price, the stated gaps and a real executed example response for each capability. It exists so a machine can decide whether DFX is worth connecting to WITHOUT connecting to it first, which is the decision most evaluating agents actually have to make.

The registry listing and this page are meant to agree. If they ever disagree, the live endpoint is the truth and this page is the thing that is wrong.