Elevation Mcp Server

Overview

Ground elevation and terrain analysis from two keyless sources: USGS 3DEP across the US and its territories (plus much of Canada and Mexico), and Open Topo Data everywhere else, which answers from SRTM GL1 v3 and, where SRTM has no tile, Mapzen terrain tiles. Look up spot heights, measure a route's ascent, descent, and grades, find an area's high and low points, and check whether terrain blocks a sightline. Runs as a stdio process or a local Streamable HTTP server.

Tools

Capability reference

elevation_get_points _tool

• points: 1–100 {lat, lon} objects in decimal degrees (WGS84)
• Each point comes back ok or no_data, with elevation_m / elevation_ft, dataset, and resolution_m (omitted for mapzen); 3DEP answers add raster_id and the upstream acquisition_date. points_with_data counts the hits, and a point without data never fails the call

elevation_get_profile _tool

• path: 2–1,000 vertices in travel order (consecutive duplicates dropped); samples: 2–250 evenly spaced points along it, endpoints included (default 100)
• summary carries total_distance_m, ascent_m / descent_m (and feet), start, end, min, and max elevation, highest_point / lowest_point, and max_grade_pct / min_grade_pct with where each occurs; ascent depends on the reported sample_interval_m. Fails as degenerate_path (under 1 m of route) or no_coverage (no sample has data)
• include_samples (default true) returns samples[] (distance, position, elevation, grade, dataset, and resolution per sample) and the sample table; false omits both, while the summary, spacing, coverage counts, datasets, notices, and attribution are unchanged, still computed from every sample

elevation_get_grid _tool

• south, west, north, east edges in decimal degrees (a box can't cross longitude 180); rows and cols 2–25 each (default 10), with rows × cols at most 250
• elevations_m and cell_datasets matrices indexed [row][col] (row 0 north, column 0 west, null without data), plus summary.highest, summary.lowest, mean_elevation_m, and relief_m. Fails as invalid_bbox, too_many_cells, or no_coverage
• Nodes are point samples with the edges included, not cell averages; re-grid a smaller box around summary.highest to refine a summit

elevation_check_line_of_sight _tool

• observer and target points, at most 1,000 km apart; observer_height_m (default 1.7) and target_height_m (default 0), each 0–10,000 m above ground; earth_model flat, geometric, optical (default, κ 0.13), or radio (κ 0.25); samples 3–250 (default 100)
• verdict is clear, blocked, or indeterminate (samples without data leave the line unconfirmed), with min_clearance_m / min_clearance_ft, the limiting_point, and the first_obstruction when blocked. Fails as same_endpoints (under 1 m apart), sightline_too_long (over 1,000 km apart), or endpoint_no_data
• Terrain only: buildings, vegetation, and Fresnel zones aren't modeled beyond what the elevation source captures; over open water, clearance is measured to the sea surface

Failures shared by every tool

A coverage miss is never an error: it degrades to no data for that point. Any provider failure fails the whole call with no partial result, and each reason's recovery text names the call to make next.

Features

Built on @cyanheads/mcp-ts-core: stdio and Streamable HTTP transports, pluggable auth (none / jwt / oauth), swappable storage (in-memory, filesystem, Supabase, Cloudflare KV/R2/D1), structured logging with optional OpenTelemetry tracing.

Elevation-specific:

• source on every tool: auto (default) queries USGS 3DEP inside its coverage and sends 3DEP misses and everything outside it to Open Topo Data; usgs_3dep and opentopodata pin one provider. Only a coverage miss falls back, never an outage
• Open Topo Data is asked for SRTM GL1 v3 first (about 30 m, land between 60°N and 56°S), then Mapzen terrain tiles where SRTM has no tile: high latitudes, Antarctica, and ocean bathymetry
• Per-call caps of 100 points or 250 samples or grid cells, inside a 45 s sampling budget; each 3DEP point is its own request, paced at 6 concurrent and 10 per second, while Open Topo Data takes 100 points per request
• The public Open Topo Data instance is paced under its published limits: one request at a time, starts at least 1.1 s apart, at most 1,000 in any trailing 24 hours per server process
• Profiles, grids, and sightlines are computed locally from point samples: haversine distances, great-circle resampling of routes and sightlines, and earth curvature with standard refraction

Agent-friendly output:

• Provenance on every value: each point, profile sample, grid cell, and sightline point names its dataset (usgs_3dep, srtm30m, mapzen), with the source's resolution_m (a resolution_m_range on profiles and grids) except for Mapzen; computed results add datasets_used counts, and a Sources: line credits every dataset that answered
• Absent stays absent: a point or sample without data omits its elevation and a grid cell is null, never 0, and computed summaries report how many values had data
• Notices flag what changes interpretation: results mixing 3DEP and Open Topo Data, Mapzen values below 0 m (sea-floor depths), sample spacing coarser or finer than the source, and thin clearance margins

Getting started

Add the following to your MCP client configuration file.

{  "mcpServers": {    "elevation-mcp-server": {      "type": "stdio",      "command": "bunx",      "args": ["@cyanheads/elevation-mcp-server@latest"],      "env": {        "MCP_TRANSPORT_TYPE": "stdio",        "MCP_LOG_LEVEL": "info"      }    }  }}

Or with npx (no Bun required):

{  "mcpServers": {    "elevation-mcp-server": {      "type": "stdio",      "command": "npx",      "args": ["-y", "@cyanheads/elevation-mcp-server@latest"],      "env": {        "MCP_TRANSPORT_TYPE": "stdio",        "MCP_LOG_LEVEL": "info"      }    }  }}

Or with Docker:

{  "mcpServers": {    "elevation-mcp-server": {      "type": "stdio",      "command": "docker",      "args": ["run", "-i", "--rm", "-e", "MCP_TRANSPORT_TYPE=stdio", "ghcr.io/cyanheads/elevation-mcp-server:latest"]    }  }}

For Streamable HTTP, set the transport and start the server:

MCP_TRANSPORT_TYPE=http MCP_HTTP_PORT=3010 bun run start:http# Server listens at http://localhost:3010/mcp

Prerequisites

• Bun v1.4.0 or higher (or Node.js v24+).
• No API key: USGS 3DEP and Open Topo Data are both keyless.
• Optional: a self-hosted Open Topo Data instance with the srtm30m and mapzen datasets, for heavy or hosted use (see Configuration).

Installation

1. Clone the repository:

git clone https://github.com/cyanheads/elevation-mcp-server.git

2. Navigate into the directory:

cd elevation-mcp-server

3. Install dependencies:

bun install

4. Configure environment:

cp .env.example .env# optionally set OPENTOPODATA_BASE_URL to a self-hosted Open Topo Data instance

Configuration

With OPENTOPODATA_BASE_URL unset, the server uses the public instance at api.opentopodata.org and paces itself under its published limits: 100 locations per request, 1 request a second, and 1,000 a day per address. Any other URL is treated as a self-hosted instance (Open Topo Data is MIT-licensed and runs in Docker) and gets 4 concurrent requests with no rate windows. That instance must serve datasets named srtm30m and mapzen and accept at least 100 locations per request; otherwise calls fail with opentopodata_config_rejected.

See .env.example for every server setting and the common framework overrides.

Known limitations

• Open Topo Data's public limits. The public instance allows 100 locations per request, 1 request per second, and 1,000 requests per day per IP address. Each server process paces itself under them, but its count is process-local and resets on restart. Several processes on one machine (a stdio server per client session), or other clients on the same address, also draw on the upstream's count, so the upstream can refuse first (opentopodata_rate_limited).
• Hosted deployments. A hosted deployment sends every user's requests from one address, so on the public instance its whole user base shares one 1,000-a-day allowance: about 1,000 point lookups, or 333 full 250-sample calls, and as few as 111 when retries spend requests too. One caller can spend that day in about 18 minutes; for up to 24 hours after, every call that needs Open Topo Data fails with opentopodata_daily_limit, including a US call in auto mode with a single 3DEP miss. A hosted deployment needs both its own Open Topo Data instance (OPENTOPODATA_BASE_URL; MIT-licensed, Docker, loaded with the srtm30m and mapzen datasets) and a per-client rate limit at its edge. Without its own instance, global answers on a hosted deployment are best-effort.
• No per-caller limits in the server. A stateless deployment without auth has no caller identity, so the server paces each upstream for the whole process and can't ration one caller against another. A hosted endpoint needs a per-client rate limit at its edge (reverse proxy, CDN, or API gateway); without one, a caller sending many heavy calls at once slows USGS 3DEP for everyone, gets other callers' heavy calls refused until its lookups drain, and can spend the public Open Topo Data day.
• Global resolution. SRTM is a 30 m radar surface model. Mapzen's 1 arc-second grid is interpolated in places from coarser sources (GMTED2010 at 7.5 arc-seconds over parts of the high latitudes, ETOPO1 at about 1.8 km over the open ocean). Outside 3DEP coverage, profile ascent is therefore underestimated, and line of sight misses terrain narrower than the source's true resolution.
• Whole-meter values. Open Topo Data's datasets are integer rasters, and the upstream rounds the interpolated result to the nearest meter, so ascent and descent over gentle terrain from Open Topo Data include 1 m quantization steps.
• Water. SRTM reports water inside a land tile as 0 m. Beyond SRTM's tiles, Mapzen reports sea-floor depth, and points, profiles, and grids include it. 3DEP values over water depend on the raster: a hydro-flattened surface, sea level, or bathymetry. Line of sight measures clearance to 0 m over Mapzen values below 0, but uses 3DEP bathymetric values as received.
• Mixed models. 3DEP is a bare-earth DEM referenced to NAVD 88 in the conterminous US (local datums in some territories). SRTM is a radar surface model that partly includes canopy and buildings, referenced to EGM96. Mapzen blends both kinds by region. Differences of a meter or more between datasets at the same point are expected; per-value provenance shows which applies.
• Sampling. Features narrower than the sample spacing are missed: a ridge between line-of-sight samples, a summit between grid nodes, short climbs between profile samples. Spacing is always reported.
• USGS 3DEP throughput. The point query service has no published limit and no batch endpoint: about 10 points per second at this server's pacing, hence the 250-sample cap and 45 s budget. Each call keeps at most 6 lookups outstanding, so concurrent calls interleave and a short call isn't stuck behind a long one. A call whose lookups would push the queue past a running call's budget is refused at once with sampling_deadline_exceeded and data.retryAfter, so two 250-sample calls at once run in turn: the first finishes, and the second can be retried after about 25 s.
• Acquisition dates from USGS are passed through raw when they have the M/D/YYYY form, and sometimes carry a zero month or day; a value in any other form is omitted.
• Antimeridian. A grid box cannot cross longitude 180, so split it into two calls. Paths and sightlines crossing it are handled by great-circle math.
• Line of sight models terrain only, with no Fresnel-zone, building, or vegetation clearance. Observer and target must be at most 1,000 km apart; within that, the curvature term is the standard parabolic approximation, which overstates the curve's rise by at most about 10 m at the midpoint.

Running the server

Local development

• Build and run the production version:

  sh

  # One-time buildbun run rebuild
  # Run the built serverbun run start:http# orbun run start:stdio

• Run checks and tests:

  sh

  bun run devcheck  # Lints, formats, type-checks, and morebun run test      # Runs the test suite

Project structure

Development guide

See CLAUDE.md for development guidelines and architectural rules. The short version:

• Handlers throw, framework catches — no try/catch in tool logic
• Use ctx.log for logging, ctx.state for storage
• Register new tools in allToolDefinitions in src/mcp-server/tools/definitions/index.ts
• Wrap external API calls: validate raw → normalize to domain type → return output schema; never fabricate missing fields

Data sources and attribution

Every successful result's Sources: line credits each dataset that answered, and a response that used any Mapzen value carries the full Mapzen attribution there. The public Open Topo Data instance publishes no terms beyond its usage limits; its server software is MIT-licensed and self-hostable.

Mapzen terrain tiles attribution, verbatim from the tilezen/joerd attribution document:

• ArcticDEM terrain data DEM(s) were created from DigitalGlobe, Inc., imagery and funded under National Science Foundation awards 1043681, 1559691, and 1542736;
• Australia terrain data © Commonwealth of Australia (Geoscience Australia) 2017;
• Austria terrain data © offene Daten Österreichs – Digitales Geländemodell (DGM) Österreich;
• Canada terrain data contains information licensed under the Open Government Licence – Canada;
• Europe terrain data produced using Copernicus data and information funded by the European Union - EU-DEM layers;
• Global ETOPO1 terrain data U.S. National Oceanic and Atmospheric Administration
• Mexico terrain data source: INEGI, Continental relief, 2016;
• New Zealand terrain data Copyright 2011 Crown copyright (c) Land Information New Zealand and the New Zealand Government (All rights reserved);
• Norway terrain data © Kartverket;
• United Kingdom terrain data © Environment Agency copyright and/or database right 2015. All rights reserved;
• United States 3DEP (formerly NED) and global GMTED2010 and SRTM terrain data courtesy of the U.S. Geological Survey.

This server is independent of the U.S. Geological Survey and of Open Topo Data, and is not endorsed by either.

Contributing

Issues are welcome. Run checks and tests before submitting:

bun run devcheckbun run test

License

This project is licensed under the Apache 2.0 License. See the LICENSE file for details.