Every tool returns a single text content block containing JSON. The examples below show the arguments object you (or the model) pass to tools/call, and the parsed JSON that comes back.

Two conventions apply everywhere:

• parcel_schema first. It is cheap, and it is how a model learns that statefp is a two-digit FIPS string and that usedesc — not usecode — holds the human-readable land use. Filters written without it tend to invent column names.
• Geometry is never returned in rows. The geom column is excluded from every row response, even if you ask for it, because raw WKT wrecks a model's context window. Use centroidx / centroidy for a point, or the spatial tools (parcel_near_point, parcel_at_point, parcel_within_bbox, parcel_within_polygon) to search by shape. The one exception is parcel_at_point, which returns the polygon on request via include_geometry. If you need polygon geometry in bulk, use the WFS endpoint directly.
• Centroids, not boundaries. parcel_near_point, parcel_within_bbox and parcel_within_polygon all test the parcel centroid, so a parcel straddling the edge of your shape is included only when its centroid is. parcel_at_point is the exception: it tests the real polygon.

Shared arguments

Argument	Applies to	Default	Notes
limit	every list-returning tool	10	Clamped to 1,000 no matter what is requested
offset	every list-returning tool	0	For pagination; must be a non-negative integer
columns	every row-returning tool	all columns except geom	Validated against the published schema; lrid is always included so results can be correlated

Passing an unknown column name, a non-integer limit, or a negative offset returns a clean error rather than a malformed query — see error shapes.

Filter operators

parcel_query and parcel_count take a where array of {column, op, value} conditions. All conditions are AND-ed together — there is no OR, and no nesting. Use in for set membership, or run two calls and combine the results.

op	Meaning	value type
eq	equals	string, number, or boolean
ne	not equals	string, number, or boolean
gt ge	greater than / greater or equal	number (or string)
lt le	less than / less or equal	number (or string)
like	case-sensitive pattern (% wildcard)	string
ilike	case-insensitive pattern (% wildcard)	string
in	member of a list	non-empty array
is_null	null test	true for IS NULL, false for IS NOT NULL

Values are type-checked and escaped before they reach the query layer, and column names are checked against the allow-list, so an invented column or a quote inside a string fails cleanly instead of producing a broken filter.

---

parcel_schema

Lists every queryable column with its type and description. No arguments.

{}

Response:

{
  "count": 53,
  "columns": [
    { "name": "lrid", "type": "uuid", "description": "Unique parcel identifier (UUID). Use this for parcel-by-id lookups." },
    { "name": "parcelid", "type": "string", "description": "Parcel ID assigned by the county assessor." },
    "…"
  ]
}

The full column set, as advertised to the model:

Column	Type	Description
lrid	uuid	Unique parcel identifier — use with parcel_by_id
parcelid	string	Parcel ID assigned by the county assessor
parcelid2	string	Alternate parcel ID
ogparcelid	string	Original parcel ID from the source record
geoid	string	Census GEOID (state + county + tract + block)
statefp	string	Two-digit state FIPS code (e.g. 06 for California)
countyfp	string	Three-digit county FIPS code
countyname	string	County name (e.g. Alameda)
taxacctnum	string	Tax account number
taxyear	integer	Tax year the assessment applies to
usecode	string	Land use code
usedesc	string	Land use description (e.g. Single Family Residential)
zoningcode	string	Zoning code
zoningdesc	string	Zoning description
numbldgs	integer	Number of buildings on the parcel
numunits	integer	Number of units (multi-family)
yearbuilt	integer	Year the primary structure was built
numfloors	float	Number of floors in the primary structure
bldgsqft	float	Total building square footage
bedrooms	float	Number of bedrooms
halfbaths	integer	Number of half bathrooms
fullbaths	integer	Number of full bathrooms
imprvalue	number	Improvement value (USD)
landvalue	number	Land value (USD)
agvalue	number	Agricultural use value (USD)
totalvalue	number	Total assessed value (USD)
taxacres	float	Tax acres
saleamt	number	Most recent sale amount (USD)
ownername	string	Owner name — indexed as full-text tokens
ownertype	string	Owner classification (government, corporate, …). Low cardinality; indexed
owneraddr	string	Owner mailing address
ownercity	string	Owner mailing city
ownerstate	string	Owner mailing state
ownerzip	string	Owner mailing ZIP
parceladdr	string	Parcel site address (street)
parcelcity	string	Parcel site city
parcelstate	string	Parcel site state (two-letter)
parcelzip	string	Parcel site ZIP
accesstype	string	Access classification. Mostly null; indexed for the non-null rows
placetype	string	Place classification. Mostly null; indexed for the non-null rows
legaldesc	string	Legal description (PLSS reference)
township	string	PLSS township
section	string	PLSS section
qtrsection	string	PLSS quarter section
range	string	PLSS range
plssdesc	string	PLSS full description
book	string	Record book reference
page	string	Record page reference
updated	datetime	Last update timestamp
centroidx	float	Parcel centroid longitude (WGS84)
centroidy	float	Parcel centroid latitude (WGS84)
surfpointx	float	Parcel surface point longitude (WGS84)
surfpointy	float	Parcel surface point latitude (WGS84)

geom is deliberately not in this list. A column a model can name is a column it can put in columns, and selecting geometry that way returns hex EWKB rather than anything readable — geometry reaches a caller as GeoJSON, from parcel_at_point or the WFS endpoint.

Only lrid, parcelid, parcelid2, ownername, geoid, ownertype, accesstype, placetype and the (centroidx, centroidy) pair are indexed on this copy. statefp and countyfp are not — filter a state or county with a geoid range instead (geoid ge '25021' AND geoid lt '25022'), or an unindexed filter over 158M rows will hit the statement timeout.

Attribute coverage varies by county — see Coverage Statistics for what is populated where, and Data Content for the full field semantics.

---

parcel_query

Run a structured filter and return matching rows. The geoid range below is Alameda County, CA — it is the indexed condition that makes the rest of the filter answerable.

{
  "where": [
    { "column": "geoid", "op": "ge", "value": "06001" },
    { "column": "geoid", "op": "lt", "value": "06002" },
    { "column": "usedesc", "op": "ilike", "value": "%single family%" },
    { "column": "yearbuilt", "op": "gt", "value": 1950 },
    { "column": "totalvalue", "op": "ge", "value": 750000 }
  ],
  "columns": ["parceladdr", "parcelcity", "ownername", "totalvalue", "yearbuilt"],
  "limit": 25,
  "offset": 0
}

Response:

{
  "type": "FeatureCollection",
  "numberMatched": 4137,
  "numberReturned": 25,
  "returned": 25,
  "features": [
    {
      "lrid": "6f1c2a94-4b3e-4f2a-9c11-0f7b2d5e8a31",
      "parceladdr": "1234 Oak St",
      "parcelcity": "Oakland",
      "ownername": "DOE JOHN A & JANE B",
      "totalvalue": 982000,
      "yearbuilt": 1962
    },
    "…"
  ]
}

features holds flat attribute objects, not GeoJSON features — there is no nested properties or geometry key. numberMatched is the total matching the filter (useful for paging); numberReturned and returned are what came back in this call.

Omit where entirely and you get an unfiltered page of the layer — occasionally useful for a sanity check, rarely what you want.

---

parcel_count

The same where filter, returning only a number. Prefer this whenever the question is "how many" — it keeps thousands of rows out of the model's context. Counts get a 90-second budget rather than the 30 seconds rows get; if one times out the total is genuinely unknown, so call again (the attempt warms the cache) or narrow the filter rather than estimating. With no filter at all, the answer is the table's planner estimate, flagged as estimated.

{
  "where": [
    { "column": "geoid", "op": "ge", "value": "51" },
    { "column": "geoid", "op": "lt", "value": "52" },
    { "column": "usedesc", "op": "ilike", "value": "%agricultur%" },
    { "column": "taxacres", "op": "ge", "value": 100 }
  ]
}

{ "count": 28914 }

---

parcel_by_id

One parcel by its lrid UUID. Non-UUID input is rejected before any query runs.

{ "lrid": "6f1c2a94-4b3e-4f2a-9c11-0f7b2d5e8a31" }

Found:

{
  "found": true,
  "parcel": {
    "lrid": "6f1c2a94-4b3e-4f2a-9c11-0f7b2d5e8a31",
    "parcelid": "1438267944",
    "parceladdr": "1234 Oak St",
    "ownername": "DOE JOHN A & JANE B",
    "totalvalue": 982000,
    "…": "…"
  }
}

Not found:

{ "found": false, "lrid": "6f1c2a94-4b3e-4f2a-9c11-0f7b2d5e8a31" }

---

parcel_near_point

Parcels whose centroid lies within a distance of a WGS84 point. The query is an indexed range scan over centroidx / centroidy, then a true-metres haversine filter on the centroid — so a large parcel whose centroid sits outside the radius is missed even when the point is inside its boundary. When you want the parcel at a coordinate, use parcel_at_point.

Argument	Required	Notes
longitude	yes	WGS84 longitude (x)
latitude	yes	WGS84 latitude (y)
distance_meters	yes	Must be positive; meters, not feet or degrees
limit / offset / columns	no	See shared arguments

{
  "longitude": -76.27836,
  "latitude": 36.87886,
  "distance_meters": 250,
  "limit": 20
}

The response is the same FeatureCollection-style envelope as parcel_query. Note the argument order: longitude first, latitude second — the same convention as GeoJSON, and the opposite of how people usually say coordinates aloud.

---

parcel_at_point

The parcel whose polygon actually contains a WGS84 point — an exact ST_Intersects test against the boundaries, not a proximity guess. This is the right tool for "what parcel is at this location".

Argument	Required	Notes
longitude	yes	WGS84 longitude (x)
latitude	yes	WGS84 latitude (y)
limit	no	Default 1. Raise it only for stacked parcels, which can genuinely overlap
include_geometry	no	Default false. true returns the full polygon — large, so ask only when you need it

{
  "longitude": -76.27836,
  "latitude": 36.87886
}

The response is not the parcel_query envelope; it is a lookup result:

{
  "found": true,
  "exact": true,
  "query": { "longitude": -76.27836, "latitude": 36.87886 },
  "parcels": [
    {
      "type": "Feature",
      "id": 7,
      "properties": { "lrid": "…", "ownername": "…", "parceladdr": "…" },
      "centroid": { "longitude": -76.2784, "latitude": 36.8789 },
      "tile": { "z": 16, "x": 18800, "y": 25585 }
    }
  ],
  "geometry_included": false
}

found: false with an empty parcels array means there is genuinely no parcel at that coordinate — the test is against real geometry, so it is not a coverage artifact of the tile archive.

---

parcel_within_bbox

Parcels whose centroid falls inside a WGS84 bounding box. This is the viewport-shaped search, and it rides the indexed (centroidx, centroidy) pair.

{
  "minx": -122.28,
  "miny": 37.80,
  "maxx": -122.26,
  "maxy": 37.82,
  "limit": 100
}

minx/maxx are longitudes, miny/maxy latitudes; each min must be less than or equal to its max or the call is rejected. Same response envelope as parcel_query. A parcel larger than the box, or one straddling its edge, is returned only when its centroid falls inside.

---

parcel_within_polygon

Parcels whose centroid falls inside a GeoJSON Polygon or MultiPolygon — city limits, a service area, a hand-drawn shape. The polygon's bounding box drives the indexed scan; each candidate centroid is then tested against the rings.

{
  "geojson": {
    "type": "Polygon",
    "coordinates": [[
      [-122.28, 37.80],
      [-122.26, 37.80],
      [-122.26, 37.82],
      [-122.28, 37.82],
      [-122.28, 37.80]
    ]]
  },
  "limit": 500
}

Rules:

• Only Polygon and MultiPolygon are accepted. A Feature or FeatureCollection wrapper is not — pass the bare geometry object (its geometry value).
• Coordinates are [longitude, latitude] in WGS84.
• Each ring needs at least four positions; an unclosed ring is closed automatically.

Same response envelope as parcel_query. Because corner rows are discarded after the fetch, a large polygon can exhaust one fetch before filling the page; when that happens the response carries a note saying results may be incomplete. For very large or complex shapes, prefer a coarse parcel_within_bbox followed by a filter, or the WFS endpoint — the 1,000-row cap applies here too.