vortx-ai/emem/emem-urban

emem-urban

Network use. The commands in this skill send JSON to https://emem.dev and receive JSON from it. https://emem.dev is the service that the MCP server of the plugin connects to. The commands do not execute what they download. This skill ships no scripts. Its commands read no environment variable, key or credential. The only local files that they read are files that an earlier step saved from https://emem.dev.

emem answers for each cell. A cell is about 10 m across. Most urban sources are much coarser than that. Thus, for each urban number, first find the area that the source measured. The band table gives that area. Step 2 shows how to go from cells to an area and give the true coverage.

The bands

emem.dev does not serve the GHSL built-surface bands. A request for ghsl.built_surface_m2 returns HTTP 400 with band_not_in_registry. That error means "this responder has no such band". It does not mean "no data here". GET https://emem.dev/v1/bands gives the full catalog. Find a band there before you build an analysis on it.

Step 1: read one place

Send this body to POST https://emem.dev/v1/recall, or call emem_recall with it:

{"cell": "defi.zb595.zbaf8.zd63c",
 "bands": ["overture.buildings.count", "overture.places.count",
           "overture.transportation.road_length_m", "overture.transportation.road_bearing_deg",
           "population", "indices.ndbi", "esa_worldcover.lc_2021",
           "surface_water.recurrence", "nightlights.dmsp_ols_avg_dn"]}

Save the answer as urban.json. Then print one line for each fact:

jq -c '.facts[] | {band, value, unit, observed_at, surface_class, args: .derivation.args, token: .memory_token}' urban.json

The cell is in Shibuya, Tokyo (35.65988 N, 139.70737 E). This call returned these values on 7 November 2026 at 01:04 UTC:

Each fact came back with an emem:fact: token in memory_token. The call fetched three bands on a miss, so the result also had a materialize_notes list. A repeat call reads the stored facts and can show no materialize_notes.

The cell reads "no buildings, no road" in the middle of Shibuya. This is correct. The cell is 10 m across and sits between features. The counts give exactly what is inside the cell. For an area question, use an area call.

Step 2: read an area

Send this body to POST https://emem.dev/v1/recall_polygon, or call emem_recall_polygon with it:

{"polygon_bbox": {"min_lat": 35.6585, "max_lat": 35.6605, "min_lng": 139.6995, "max_lng": 139.7015},
 "bands": ["overture.transportation.road_length_m", "overture.buildings.count"],
 "budget_ms": 60000}

Save the answer as area.json. Then print the coverage and the sums:

jq '{area_km2, cells_sampled, native_grid_cells_in_polygon, coverage_fraction, is_exhaustive}' area.json
jq -c '[.by_cell | .. | objects | select(.band? == "overture.transportation.road_length_m") | .value]
  | {n: length, nonzero: (map(select(. > 0)) | length), sum: add}' area.json
jq -c '[.by_cell | .. | objects | select(.band? == "overture.buildings.count") | .value]
  | {n: length, sum: add}' area.json

emem_recall_polygon is not in the short tool list that the MCP server of the plugin shows. The server runs it by name through tools/call. If your host calls only listed tools, use the REST route. An MCP result has a wire budget of about 24 KB. A larger result comes back slimmed, and _emem_truncation names what it left out. The REST route returns the full response. On 7 November 2026 it was about 400 KB for this box.

This call returned these values on 7 November 2026 at 01:06 UTC:

The 126.5 m is a sample. It is not the road length of the box. An estimate from the mean for each cell is 126.5 / 64 x 400, about 790 m. That is about 20 km of carriageway for each km2. When you give this number, say that it is an estimate from a 16 percent sample. Cite the facts that it came from.

Do not scale the building count in the same way without a warning. A centroid count in 10 m cells is sparse. A count of 3 in 64 cells has a wide interval. To get street grids and building totals over a full district, sample more cells with cells_per_sqkm. Or list the full grid with POST /v1/cells_in_bbox and read it with POST /v1/recall_many (at most 256 cells for each call).

Read the evidence correctly

• Overture facts name their release in sources, with a hash of the

file listing of that release. Thus a figure stays checkable after Overture removes the release.

• population, nightlights.dmsp_ols_avg_dn and

esa_worldcover.lc_2021 give the value of the source pixel at the cell. Two adjacent cells can be the same pixel. Do not subtract them and call the result a gradient. The spatial_basis object in the response gives the same warning.

• The data years are different. Population is 2020, WorldCover is

2021, DMSP is 2013, and Overture is a 2026 release. Give each date. Do not give one "as of" date for all of them.

• The built-up signals can disagree for real reasons. NDBI can read

bare soil as roofs. WorldCover is the output of a 2021 classifier. Overture gives mapped footprints. Say which signal you used.

To cite a figure in a report, see `emem-research-grade-citation`.