Use the data without flattening its meaning.

01 / Release status

Production data plane live; payment activation remains.

The 21-table data model, PostGIS migration, replay-safe 521-event import, five-tool HTTP and MCP runtime, x402 discovery, machine manifests, and registry packages are implemented. https://api.businesschanges.dev is live with 521 imported events. x402 enforcement remains disabled until the recipient wallet and first reconciled settlement are approved.

02 / Discovery

Machine-readable before marketplace-specific.

03 / Query translation

The calling agent translates language.

The API does not invoke a model or accept an unbounded free-text query. The calling agent maps user intent into strict typed arguments; the database path is deterministic and auditable.

{
  "jurisdictions": [
    "US-FL"
  ],
  "geography": {
    "cities": [
      "Orlando"
    ]
  },
  "intents": [
    "opening_leads"
  ],
  "industries": [
    "restaurant"
  ],
  "observedAfter": "2026-07-01T00:00:00.000Z",
  "projection": "compact",
  "limit": 25
}

04 / Tool sequence

Search compact, then buy detail selectively.

1. Inspect free coverage when geography or absence matters.
2. Call find_business_change_leads with projection: compact.
3. Rank compact events in the calling agent.
4. Expand only selected IDs with get_business_event_evidence.
5. Save the observed-time cursor for recurring monitoring.

05 / Data semantics

Time, identity, and absence stay separate.

• effectiveAt is when the event happened.
• firstObservedAt is when the platform discovered it.
• An exact premises and its parent building are distinct match scopes.
• Classification, entity resolution, and public status have separate confidence.
• Regulatory activity never proves public opening or closure.
• No-match is meaningful only inside the returned source coverage and window.
• Superseded and retracted events remain visible in the delta feed.

06 / Quality gate

High-confidence signals by default.

Paid search and delta calls default to strong evidence and classification confidence of at least 0.75. The current commercial-default cohort contains 445 of 521 events. A 130-record stratified audit verified every official-source classification chain, with zero contradictions and zero public-opening overclaims.

Independent entity corroboration and external event corroboration are reported separately. Source fidelity does not silently become a claim that a location is open to the public.

07 / MCP

One stateless remote server.

The server uses Streamable HTTP, lists five typed read-only tools, validates structured output, and needs no sticky session. The same URL can be installed directly in Claude, Cursor, Codex, and generic MCP clients.

{
  "mcpServers": {
    "business-changes": {
      "type": "http",
      "url": "https://api.businesschanges.dev/mcp"
    }
  }
}

08 / HTTP

Strict POST JSON.

POST /v1/events/search
Content-Type: application/json

{
  "jurisdictions": [
    "US-FL"
  ],
  "geography": {
    "cities": [
      "Orlando"
    ]
  },
  "intents": [
    "opening_leads"
  ],
  "industries": [
    "restaurant"
  ],
  "observedAfter": "2026-07-01T00:00:00.000Z",
  "projection": "compact",
  "limit": 25
}

Unknown fields are rejected. Results include versioned metadata and coverage where needed. Keyset cursors are opaque and tied to either effective-time or observed-time ordering.

09 / Distribution

Canonical direct surface, many discovery layers.

The public MCP and OpenAPI endpoints remain the product. The official MCP Registry and Smithery index that endpoint; OpenAI and Cursor package it as a plugin; Claude accepts it directly; Monid and BlockRun are provider channels; x402 Bazaar advertises paid HTTP resources.

Registry presence is not success by itself. We measure name-blind tool selection, valid calls, evidence expansion, repeat delta usage, and paid conversion by channel.

10 / Errors

Correct, retry, or qualify.