OpenContext

San Diego City GIS MCP — a City of San Diego fork of OpenContext (forked from the Worcester GIS fork). It serves the City's ArcGIS Server REST services directory (webmaps.sandiego.gov/arcgis/rest/services, ArcGIS Enterprise 11.5) through the built-in arcgis plugin.

It is a sibling to the San Diego regional (SANDAG/SanGIS) and Worcester servers: identical tool names and signatures, so it composes with them at the MCP client with zero new orchestration. The regional server covers county-wide layers; this one covers City-authored layers (MHPA, Base Zones, Community Plan Land Use, and ~700 more).

How discovery works (no Hub here)

webmaps.sandiego.gov is a bare services directory — there is no ArcGIS Hub / Open Data catalog in front of it, so the Hub-search discovery used by the Worcester fork does not apply. Instead:

1. scripts/crawl_catalog.py walks the directory offline: folders → MapServer/FeatureServer services → each service's /layers?f=json, capturing layer id, name, geometry type, description, extent, and maxRecordCount.

2. The result is serialized to plugins/arcgis/catalog.json — a versioned, diffable deploy artifact (708 layers from 327 services at the 2026-09-09 crawl).

3. The running server loads that manifest at startup — instant, no live crawl, no cold-start penalty. search_datasets does substring/fuzzy/acronym matching over it.

4. Services that require an ArcGIS account (HTTP 401/403 or ArcGIS error codes 498/499) are skipped during the crawl and recorded in the manifest's skipped list — 18 services on this host at last crawl (AMPGIS, GetItDone, TED, the PUD_* services, four legacy aerials, …).

To refresh the catalog: python scripts/crawl_catalog.py, review the diff, commit, redeploy.

Related MCP server: MCP Data Vermont

Dataset IDs

A dataset id is the layer's services-directory path:

{folder}/{service}/{MapServer|FeatureServer}/{layerId}

Known-good public layers (all in the anonymous Planning/PLN_LongRangePlanning MapServer):

These are curated as featured_datasets in config.yaml (aliases + notes boost search; edits take effect on deploy without re-crawling).

The WGS84 contract

The layers are authored in EPSG:2230 (NAD83 State Plane California Zone VI, US survey feet). Every query this server sends sets inSR=4326 and outSR=4326, so all tools take and return WGS84 lon/lat — same as the sibling servers. Without inSR, WGS84 coordinates would be interpreted as State Plane feet and silently return zero rows.

Pagination is metadata-driven: each layer's maxRecordCount comes from the catalog (it varies by layer), and query_data pages with resultOffset/resultRecordCount when needed.

Connect to the server

Add it as a custom connector in Claude (same steps on Claude.ai and Claude Desktop):

1. Settings → Connectors (or Customize → Connectors on claude.ai)

2. Add custom connector

3. Name it e.g. San Diego City GIS and paste the URL:

   https://sandiego-city-gis.codeforanchorage.org/mcp

Quick health check from a terminal:

curl -sS -X POST https://sandiego-city-gis.codeforanchorage.org/mcp \
  -H 'Content-Type: application/json' \
  -d '{"jsonrpc":"2.0","id":1,"method":"ping"}'
# → {"jsonrpc":"2.0","id":1,"result":{}}

Tools exposed

Raw JSON-RPC example:

curl -sS -X POST https://sandiego-city-gis.codeforanchorage.org/mcp \
  -H 'Content-Type: application/json' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call",
       "params":{"name":"arcgis__search_datasets",
                 "arguments":{"q":"MHPA","limit":5}}}'

Verified end-to-end

The definition-of-done check is a WGS84 point-in-polygon on MHPA at the Tijuana River Valley (32.5539, -117.0846) — a point that returns null on the regional server's County MSCP_CN layer but sits inside a City MHPA preserve:

// arcgis__spatial_query_point
{
  "item_id": "Planning/PLN_LongRangePlanning/MapServer/7",
  "lon": -117.0846,
  "lat": 32.5539,
  "out_fields": "HABPRES,INHABPRES,SUBAREA,ACRES"
}

returns the containing preserve polygon:

Record 1:
  HABPRES: 100
  INHABPRES: Yes
  SUBAREA: 113
  ACRES: 2734.79

scripts/smoke_prod.py runs this plus the search ("MHPA", "zoning"), schema, TOTAL MATCHING, geocode → zoning chain at City Hall, structured-output, and MCP conformance checks against any deployment:

python scripts/smoke_prod.py                             # production
python scripts/smoke_prod.py http://localhost:8000/mcp   # local

Local development

uv sync                              # or: pip install -r requirements.txt
python scripts/crawl_catalog.py      # (re)build plugins/arcgis/catalog.json
python scripts/local_server.py       # serves http://localhost:8000/mcp
python -m pytest tests/ -q           # tests

See CLAUDE.md and docs/ for architecture, deployment, and plugin development.