Skip to main content
Glama
sar-geo-mcp

sar-geo-mcp

by paulzcgu

sar-geo-mcp

Search and rescue planning tools for Claude, built on CalTopo maps.

SAR Geo is a ready-made set of tools, features and scenarios that a language model can call directly. Without it, a model asked to plan a search has to reinvent the work every time - writing its own coordinate parsing, slope arithmetic or route search from scratch, and getting a slightly different answer each time. With it, the model calls one tested tool and gets the same measured answer for the same question. That makes results stable, and makes them adjustable: change a tool's parameters, not the model's improvised code. It is also a plain Python package, so any Python program can import the same functions as a library, with no MCP server and no language model involved.

SAR Geo gives Claude forty-seven tools for planning a search: read a CalTopo map, stage the terrain around a place last seen, and ask for hasty search lines, parking, helicopter landing zones, the way out, walking-time rings, the nearest hospitals and trauma centers, and urban assignments of one size. Four scenarios run several of those in one call: a medevac, the initial response, an urban search and an assignment debrief. Results come back as facts with GeoJSON files CalTopo imports as draft assignments.

It is read-only: it never writes to a CalTopo map. The tools measure; the people running the search decide.

Install

Claude Desktop (no coding needed)

  1. Download sar-geo.mcpb from the latest release (direct link: sar-geo.mcpb).

  2. Open it, and click Install in Claude Desktop (macOS or Windows). Claude warns that the extension is not verified by Anthropic; that is expected for anything not in Anthropic's directory.

  3. Fill in any keys you have (all optional; see Keys below) and save.

The first start downloads the server from PyPI, which takes a minute or so; later starts reuse that copy, and offline it runs the copy it already has. It updates itself: each time Claude Desktop starts it, it runs the newest release, so you install the .mcpb once. Ask Claude "check for a sar-geo update" to see your version, or set Pin version to stay on one, such as during an operation.

Any other MCP client (Claude Code, and others)

The package is on PyPI and needs Python 3.12 or newer. The mcp extra adds the server; without it you get the library alone. With uv, nothing needs installing first:

claude mcp add sar-geo -- uvx --python 3.12 --from "sar-geo-mcp[mcp]" sar-geo-mcp

Or install it with pip and register the command it puts on your PATH:

pip install "sar-geo-mcp[mcp]"
claude mcp add sar-geo -- sar-geo-mcp

Other clients take the same command (sar-geo-mcp) as a stdio server.

From source

git clone https://github.com/paulzcgu/sar-geo-mcp.git
cd sar-geo-mcp && pip install -e ".[mcp]"

AGENTS.md covers the architecture, tests and releases.

Related MCP server: mcp-server-geo-optimizer

Use it from Python

Every tool is also a plain Python function, and the modules behind them can be imported on their own. A script, a notebook or another program gets exactly what an MCP caller gets:

from sar_geo_mcp.coords import parse_location

parse_location("10S EG 61234 41234")          # (37.4158..., -122.3080...)

# A dropped western minus is put back only against a reference point:
parse_location("37.3, 122.32", near=(37.3045, -122.32))   # (37.3, -122.32)

A tool function takes a ServerContext first, then the tool's own arguments:

from sar_geo_mcp.config import CaltopoConfig
from sar_geo_mcp.context import ServerContext
from sar_geo_mcp.components.coords import distance_bearing

ctx = ServerContext.build(CaltopoConfig.from_env())
distance_bearing(ctx, 37.3, -122.32, 37.3045, -122.32)
# about 499 metres, bearing 0 (due north)

Install with pip install sar-geo-mcp (no [mcp] needed). Features live in sar_geo_mcp.features, scenarios in sar_geo_mcp.scenarios; AGENTS.md lists which module holds each tool.

Keys

Every key is optional. Without any, public CalTopo maps, terrain, OpenStreetMap, parking, hasty lines, egress and hospitals all work. Ask Claude "what's my sar-geo key status?" to see which are set and what each unlocks.

Key

Turns on

Where to get it

CalTopo WMS token

Terrain pictures, slope shading, the canopy panel, landing-zone imagery, urban blocks

CalTopo → your name → Your Account → API Access: the six characters between /api/ and /wms (paid plan or team)

CalTopo credential ID + key

Private and team maps

CalTopo → Your Account → Credentials (a read-only team credential is enough)

Mapbox token

Mapbox satellite pictures of landing zones

account.mapbox.com

Google Maps key

Google satellite pictures of landing zones

Google Cloud console, Maps Static API (its terms limit how images may be used)

From source, set them as ZA_WMS_TOKEN, ZA_CALTOPO_ID and ZA_CALTOPO_KEY (or a cts.ini), ZA_MAPBOX_TOKEN and ZA_GOOGLE_MAPS_KEY. The CalTopo account ID field in the extension is only for printing PDFs through CalTopo; no tool needs it today.

Where files go

Ask Claude to save a file (an export, an imagery chip) and it goes in Documents/sar_geo/<chat folder>, one folder per chat, and opens in Finder or Explorer. Downloaded terrain is cached out of the way, in ~/Library/Caches/sar_geo on a Mac and %LOCALAPPDATA%\sar_geo\cache on Windows (~/.cache/sar_geo elsewhere), and is downloaded again if deleted. It can grow to a few GB.

Both folders can be changed: in the extension's settings (Exports folder, Cache folder), or from source with SAR_GEO_EXPORTS_DIR and ZA_CACHE_DIR.

The tools

Each tool carries its layer as an MCP tag - component, feature or scenario - so a client can list one layer; AGENTS.md says which tool is in which. The four scenarios, last in the table, call several features at once and answer in compact sections, each with its own status and the exact feature call that returns the full answer.

Tool

What it answers

list_assignments

What assignments are on this map, and where

get_assignment

One assignment in full — geometry, description, centre

list_map_features

Everything on the map, filterable by class

parse_coordinate

Read a coordinate a person wrote or said

convert_coordinate

The same point in DD, DDM, DMS, UTM, and MGRS/USNG

distance_bearing

Geodesic ground distance and initial bearing

point_in_geometry

Is this clue inside the assignment, and how far from the edge

geometry_bounds

Bounding box and centre of any GeoJSON geometry

boundary_clearance

How much of a boundary's inside surrounds this, and whether the way in stays inside it

buffer_geometry

A line widened into the ribbon CalTopo's Buffer draws, so much each side; single_ring cuts a loop's hole open so CalTopo keeps it

prefetch_map

Download a map's ground now, before losing connectivity

cache_status

What the tile cache holds, by tier

evict_cache

Trim the cache to a ceiling, oldest maps first

save_export

Copy a file a tool wrote (an export GeoJSON, a chip) to ~/Documents/sar_geo/<chat folder> and show it in Finder or Explorer

key_status

Which optional keys are set (never their values), what each unlocks and where to get it

check_for_update

The version running against the newest on PyPI, and how to update

list_data_sources

Every external data source the server reads - what it is, its coverage and attribution, whether its key is set (never the key), and which tool pulls it

fetch_raster

One raster source's own image of an area, saved as a file with a short summary: elevation, vegetation type, any CalTopo map layer, NAIP imagery or Mapbox satellite. Raw data, without the planning tools' checks

query_features

One feature source's own features over an area as GeoJSON: OpenStreetMap ways and nodes filtered by tag, the land agencies' road and trail records, or hospitals. Raw data, without the planning tools' checks

prepare_terrain

Stage one assignment's ground; returns the handle everything below names

terrain_metrics

Everything measured for a staged assignment — the debrief bundle

identify_features

Just the feature kinds you name — steep bands, drainages, clearings, cliff bands, cliff_pitch (ranked by the drop you would rig on), routes, ingress

list_feature_kinds

Which feature kinds identify_features can look for, with their assumptions

slope_stats

Slope statistics inside one named shape — "quantify that sidewall"

slope_shading

CalTopo's slope bands beside this package's own, and the patches where the DEM reads shallower

shaded_relief

Terrain to look at, and benches OpenStreetMap has not got - clean and OSM-overlaid, plus measured lineaments

unmarked_paths

Roads and trails OpenStreetMap has not got, from Forest Service, BLM and USGS inventories - closed Forest Service roads included - each checked against NAIP imagery for a corridor still bare on the ground, ranked by how many independent inventories saw it, junction gaps bridged along flat ground, and written to a GeoJSON file CalTopo imports as coloured lines. DEM benches are an experimental source, read only when asked for

elevation_profile

Elevation along a drainage centreline — how steep is the channel

coverage_for

Coverage and gaps for a staged assignment, from named GPS tracks

prepare_area

Stage the ground around a place last seen (point + radius, in m, km, ft, yd or mi) or a polygon; returns an aoi- handle that never goes stale, and the area's outline to draw

assignment_norms

What real assignments look like, from a corpus of past searches, with provenance; per-resource target sizes

site_facts

Can a team occupy and work this ground - drop, span, the bench above, the landing below, and cover on all three

describe_boundary

What an assignment's or drawn shape's edges follow, how it was drawn, what it holds, against corpus norms

hasty_candidates

Day-1 hasty candidates around a staged area: trails and roads as lines by default, mapped waterways as corridors when include names them (no drainages or ridges), measured and ordered by distance

parking_candidates

Where a search can park near a staged area: public lots and public hardstand, nearest first, each with a point guaranteed to sit inside it and, since OSM rarely names a car park, the nearest named trail or road within 150 m

lz_candidates

Where a helicopter can put down near a staged area: open, flat patches graded preferred, minimal or marginal per the CalTopo safety course, with a touchdown point, the nearest power line or "unknown", the road it touches or sits beside, the lot it is in, and an imagery chip to look at

egress_routes

The way out on the ground from a location: the quickest exit a 2WD vehicle can reach and the quickest a 4x4 can, on foot or by litter, each with its route, minutes, distance, climb and descent, and the road, lot and gates it ends on

travel_time_rings

Rings of equal walking time round a place last seen - the travel-time counterpart of a radius circle - stretched along trails, pinched by cliffs and unbridged rivers, slowed by creek crossings and brush

nearest_medical_facilities

The closest medical center, trauma center and pediatric trauma center to a point, ranked by road time, as a table: trauma level, helipad Y/N, road distance and time, air time where there is a helipad, phone and address - plus markers to draw. Stateless

populated_areas

Urban and suburban search blocks: the built-up lots along each street, both sides and the street between them, outlined as one block per street from CalTopo's parcel and structures layers; apartment complexes, schools and shopping centres as blocks of their own

urban_assignments

Those blocks merged into assignments of about one size (the corpus median for the resource, or a target you give), never across an arterial

caltopo_export

Any candidate set as a GeoJSON file CalTopo imports as DRAFT assignments, with the search area and house markers - a file, never a write to CalTopo

get_candidate

One candidate in full, geometry included, by id

plan_medevac

A patient at a point, in one call: the landing zones nearest them, the 2WD and 4x4 exits by litter and on foot, and the nearest hospitals and trauma centers - compact sections, each with its own status and the exact call for the full answer, plus one GeoJSON file CalTopo imports. Facts, no verdict

plan_initial_response

The first operational period from a place last seen, in one call: the hasty lines a team can run (trails and roads by default; waterways through overrides), the nearest parking, walking-time rings, the landing zones nearest the place and the nearest hospitals - compact sections, each with its own status and the exact call for the full answer, plus the hasty lines as DRAFT CalTopo assignments and a second GeoJSON file of everything else. Facts, no verdict

plan_urban_search

An urban or suburban search round a place last seen, in one call: the populated blocks grouped into assignments of about one size, the nearest parking and the nearest hospitals - compact sections, each with its own status and the exact call for the full answer, plus the assignments as DRAFT CalTopo assignments and a second GeoJSON file of parking and hospitals. Needs the CalTopo WMS token for the blocks. Facts, no verdict

debrief_assignment

A team back from a CalTopo assignment, in one call: what the ground was (slope, landform, drainages, vegetation, sweep width, ways in), how much of it the team's GPS tracks covered and where the gaps are - each gap's size, the terrain it leaves unsearched and a point inside it - and the terrain features worth asking about - compact sections, each with its own status and the exact call for the full answer, and the boundary check said once. Facts, no verdict

parse_coordinate handles decimal degrees, degrees-decimal-minutes, DMS, UTM, and MGRS/USNG, and it refuses to guess: a transposed pair or a dropped western minus is resolved only against a reference point you supply, and a grid reference that decodes far from that point comes back unparsed rather than plausible. A wrong coordinate is worse than no coordinate.

Documentation

Each tool describes itself to Claude, so asking Claude how to use one is the quickest guide. AGENTS.md has the architecture and how to change, test and release the code.

Data

Elevation (USGS 3DEP), vegetation type (LANDFIRE), roads and trails (OpenStreetMap, Forest Service, BLM, USGS), hospitals (HIFLD), road routing (OSRM), state boundaries (Census) and NAIP imagery are public. CalTopo's map layers need the WMS token above; Mapbox and Google satellite images need their own keys. Ask Claude to "list the sar-geo data sources" for each one's coverage and attribution. Coverage is the United States.

Origin

Extracted from a private SAR debrief-automation project, which is now its first consumer.

License

MIT - see LICENSE. It comes with no warranty: SAR Geo measures ground and reports facts, and every decision in a search stays with the people running it.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      D
      maintenance
      Enables AI assistants to perform geospatial analysis, spatial data processing, vector and raster operations, coordinate transformations, spatial statistics, mapping, and access to climate, satellite, and other geospatial data through MCP-compatible clients.
      MIT
    • A
      license
      Not graded
      quality
      B
      maintenance
      Enables agents to retrieve ground elevations worldwide, profile route ascent/descent and grades, sample grid areas for high/low points and relief, and test whether terrain blocks a sightline, drawing on USGS 3DEP and Open Topo Data with per-value provenance.
      351 npm
      3
      Apache 2.0
    • A
      license
      A
      quality
      B
      maintenance
      Enables Claude to generate mountain bike loops or point-to-point routes from a place name or coordinates, computed on real OpenStreetMap trails by a routing engine and saved as GPX files with elevation, along with a summary and warnings about traffic, steep grades, protected areas and forbidden access. It also lists previously saved routes.
      2
      MIT