Wherobots MCP Server: Building GEOINT Spatial Pipelines with AI Agents

Editor’s note: The Wherobots Spatial Data Catalog is now the Havasu Catalog.

I built three national-security GEOINT use cases on the Wherobots stack in days instead of weeks. A Critical Infrastructure Vulnerability (CIV) pipeline with two regional variants, plus a border-corridor analysis on real transportation segments. The Wherobots geospatial MCP server is what made that timeline possible.

Most of the work in standing up a credible use case isn’t the analysis. It’s the plumbing. Finding the right data within the organization. Understanding the schema. Anchoring synthetic operational data to real geometries so the joins represent reality. That’s where the weeks go, and anyone who’s built a multi-table spatial pipeline on real data knows the shape of it.

About a dozen joined results, multiple Overture Maps Foundation themes, notional mission overlays, and an scenario-based MCP server evaluation notebook driving the whole thing. The build took days instead of weeks thanks to the Wherobots Spatial AI Coding Tools.

What used to be the slow part of this kind of build was never the analysis. It was the getting to the analysis. Finding the right Overture tables. Learning their column shapes. Remembering whether class lived on base_infrastructure or places_place. Figuring out that categories was structured as a STRUCT vs. a flat field. Discovering the hard way that power_pole and power_tower collectively contain 36M rows and could have consumed tons of unnecessary resources. The Wherobots MCP collapsed every one of those build time consuming events into a tool call.

Designing the CIV Pipeline Schema

The end state for the CIV use case is a 10-step pipeline ending in a 5-component scoring formula:

vulnerability_score = 0.25 * flood_count
                    + 0.25 * wildfire_count
                    + 0.15 * (facility_count + 0.01 * building_count)
                    + 0.20 * criticality_tier * 100
                    + 0.15 * (negative_tweet_count * 2 + neutral_tweet_count * 0.5)

Sitting underneath that score: an AOI layer, an asset layer, flood and wildfire proximity buffers, settlement/facility proximity, building density, a criticality tier table, seasonal hazard windows, social-media sentiment proximity, and a geofence overlay. About a dozen tables in wherobots.geoint.*_civ, each with a clear FK relationship back to critical_assets_civ.asset_id. Overture thematic geometries underneath, synthetic operational signal layered on top.

That’s a lot of schema. Designing it is the interesting part. Creating the plumbing without agentic support can extend the timeline significantly.

Direct Access to Your Spatial Catalog

The real paradigm shift isn’t any single tool call. It’s that the agent can reach for my Wherobots organization’s current catalog, and any of the catalogs connections it has (Databricks Unity, AWS Glue, S3) whenever it needs to, intra-session. Not a snapshot from when the analysis was developed. Not a schema description I pasted into a system prompt three turns ago. When the agent invokes the MCP, the response reflects the catalog as it exists at that moment, which means:

• After I created critical_assets_civ in Step 2, a follow-up prompt could find it via a fresh catalog walk. No need to tell the agent anything.
• When I added a negative_tweet_count column further down the pipeline, the next describe_table call against that table reflected the new column.
• When I joined two of my own tables and persisted the result as a new Iceberg table, the agent could discover and reason about it on a subsequent prompt as easily as any Overture table.

The catalog and the agent’s understanding don’t have to drift apart. That sounds like a small thing. In practice though, it is a BIG deal, because the work of building a multi-table use case is a sequence of “I just created this, now use it to build the next thing.” Whenever the agent reaches for the catalog, it can see what I’ve built so far, what’s in the open data catalog, and what fields are queryable. I’m never re-explaining my own schema.

This is what an AI context engine for the physical world looks like in practice. The agent isn’t reasoning about geography from training data. It’s reading the live state of a interconnected and extensible Wherobots spatial catalog and writing SQL grounded in it.

The loop that emerged on top of that capability looked the same across both builds:

1. Discover. “What Overture tables describe critical infrastructure in this AOI?” The MCP catalog walk surfaced base_infrastructure, base_water, base_land_cover, places_place, and buildings_building, including the row counts that warned me off power_pole and power_tower before I computed a spatial join on them.

-- Catalog walk: filter base_infrastructure to classes worth joining on,
-- excluding power_pole (18.3M rows) and power_tower (18.0M rows) once
-- list_tables_tool surfaced their volume.
SELECT class, COUNT(*) AS row_count
FROM wherobots_open_data.overture_maps_foundation.base_infrastructure
GROUP BY class
ORDER BY row_count DESC;

2. Describe. Before any SQL got written, describe_table returned the actual columns. No guessing whether categories.primary was a STRUCT field or a flat string. No discovering at runtime that subtype and class were different axes.
3. Generate with grounded context. Schema-aware SQL generation produced queries that ran on the first or second attempt because the agent had already seen the column shape, including the tables I’d just created moments earlier.

-- Step 3: Flood Hazard Proximity
-- Counts water features within 1 km of each critical infrastructure asset.
-- Pre-filters water to Ukraine AOI to reduce spatial join volume.
SELECT
  a.asset_id,
  COUNT(w.id) AS flood_water_count
FROM critical_assets a
JOIN wherobots_open_data.overture_maps_foundation.base_water w
  ON ST_DWithin(a.geometry, w.geometry, 1000, true)
CROSS JOIN ukraine_aoi ua
WHERE w.class IN ('river', 'stream', 'lake', 'reservoir', 'canal')
  AND ST_Intersects(ua.geometry, w.geometry)
GROUP BY a.asset_id
ORDER BY flood_water_count DESC;

4. Execute with limit=10. Every query got tested on a small slice before it ran across the full AOI. The cost of being wrong dropped to near zero, which meant I iterated on scoring weights, distance thresholds, and asset-class filters far more aggressively than I would have otherwise.

The result: I spent my time on the parts that required judgment. Where should the flood buffer be: 500m, 1km, 2km? 500m is too tight in dense urban terrain. 2km swallows half the AOI. The MCP let me test all three in the time it would normally take to write the first query. What’s the right tier-weighting for a substation vs. a generator? Does negative sentiment matter 2x or 4x more than neutral sentiment in the composite score? How do I express a seasonal vulnerability window so the downstream MCP scenarios can filter on it cleanly?

Here’s what the scoring formula above looks like as the actual SQL that runs against the persisted views. Every *_proximity table shares the same asset_id grain, so the composite score is a single set of LEFT JOINs back to the asset layer:

-- Step 10: Enhanced Vulnerability Score (5-component composite)
SELECT
  ca.asset_id,
  ca.class,
  ac.criticality_tier,
  COALESCE(fp.flood_water_count, 0)        AS flood_count,
  COALESCE(wp.wildfire_landcover_count, 0) AS wildfire_count,
  COALESCE(sp.facility_count, 0)           AS facility_count,
  COALESCE(bd.building_count, 0)           AS building_count,
  COALESCE(sm.negative_tweet_count, 0)     AS negative_tweet_count,
  (
    0.25 * COALESCE(fp.flood_water_count, 0) +
    0.25 * COALESCE(wp.wildfire_landcover_count, 0) +
    0.15 * (COALESCE(sp.facility_count, 0) + 0.01 * COALESCE(bd.building_count, 0)) +
    0.20 * ac.criticality_tier * 100 +
    0.15 * (COALESCE(sm.negative_tweet_count, 0) * 2
          + COALESCE(sm.neutral_tweet_count, 0) * 0.5)
  ) AS vulnerability_score,
  CASE
    WHEN COALESCE(fp.flood_water_count, 0) > 0
     AND COALESCE(wp.wildfire_landcover_count, 0) > 0
    THEN TRUE ELSE FALSE
  END AS multi_hazard_flag
FROM critical_assets ca
JOIN      asset_criticality      ac ON ca.asset_id = ac.asset_id
LEFT JOIN flood_proximity        fp ON ca.asset_id = fp.asset_id
LEFT JOIN wildfire_proximity     wp ON ca.asset_id = wp.asset_id
LEFT JOIN settlement_proximity   sp ON ca.asset_id = sp.asset_id
LEFT JOIN building_density       bd ON ca.asset_id = bd.asset_id
LEFT JOIN social_media_proximity sm ON ca.asset_id = sm.asset_id
ORDER BY vulnerability_score DESC;

The Same Pattern Works for Retail, Insurance, and Supply Chain

Strip the national-security framing off the build and the underlying loop is domain-agnostic. Swap critical_assets_civ for store locations and the proximity layers for foot-traffic decay and competitor density, and the same pipeline becomes a retail expansion analysis. Swap them for policy-in-force locations and named-storm tracks, and it’s catastrophe exposure modeling for an insurance carrier. Swap them for distribution centers and port-congestion signal, and it’s a supply-chain resilience build. Swap them for parcels and zoning and hazard layers, and it’s a municipal planning workflow.

The paradigm generalizes: a live catalog the agent re-reads every turn if needed, layered data anchored to real datasets and foundational Overture data, and an iterative discover → describe → generate → test → refine loop. It works anywhere you have a catalog of geospatial or tabular data you’re actively building on top of. The schema work is the same shape regardless of whether the bottom layer is Overture infrastructure, OpenStreetMap, NAIP rasters, parcel records, or a customer’s own warehoused operational data. What changes is the vocabulary on top.

What You Ship: Notebooks, Schemas, and a Reusable GEOINT Pipeline

Two regional CIV setup notebooks, a third for border-corridor segments, and an MCP scenario notebook with 10+ analyst prompts that exercise the full schema. All on real Overture geometry. All synthetic where it needs to be. All persisted to our managed Iceberg catalog as wherobots.geoint.* so the next analyst can pick it up where I left off.

The deliverable looks like weeks of work. The MCP server and spatial AI coding tools made it days. The schema is where the value lives, and the AI coding tools let me spend my time there.

Here is a link to the use case setup notebook

Get Started with the Spatial AI Coding Assistant

Access Now