@mailwoman/mcp

An MCP server exposing Mailwoman's parse/geocode/POI toolset to agents over stdio — no HTTP endpoint, just a subprocess an MCP client launches.

Tools

| Tool | What it does | Needs the model + a gazetteer | | -------------------------------- | ---------------------------------------------------------------------------- | ----------------------------- | | mailwoman_parse | Runtime-pipeline parse (optionally POI-aware) | Yes | | mailwoman_geocode | Street-level geocode cascade | Yes | | mailwoman_poi_search | POI-intent extraction, executed against a wired poi.db | Yes | | mailwoman_overpass_export | Renders a POI query as OverpassQL (prints the query, never runs it) | Yes | | mailwoman_layer_manifest | Reads a spatial-layer database's provenance manifest + coverage summary | No | | mailwoman_bdc_filing_landscape | Reads a bdc.db layer's filing census over census blocks or H3 cells | No | | mailwoman_plausibility_check | Scores one claimed broadband-service assertion against filing + POI evidence | Only when it geocodes | | mailwoman_filer_lookup | Reads the FCC filer identity crosswalk from a filer.db layer | No | | mailwoman_filer_family | Reads a corporate family's membership from a filer.db layer | No |

The four model-backed tools load the en-US weights and open a resolver on the first call that needs them, not at startup. With neither $MAILWOMAN_CANDIDATE_DB nor a WOF distribution on the data root, that first call answers with the mailwoman data pull candidate fix instead of an internal resolver error.

Config

{
	"mcpServers": {
		"mailwoman": {
			"command": "mailwoman-mcp",
			"args": ["--poi-db", "/path/to/poi.db"]
		}
	}
}

--poi-db <path> wires mailwoman_poi_search (and mailwoman_parse's poi: true path) to a real database; omit it and those tools degrade gracefully to intent-only.