groundstation

Earth data, agent-ready. A Development Seed labs prototype that puts the cloud-native geospatial stack in the hands of AI agents — and of anyone with a browser.

Ask a question about any place on Earth. groundstation finds the freshest satellite imagery, active fires and disaster alerts, and the weather, does the pixel math, and hands back an interactive map — through whichever door fits:

flowchart LR
    A["🗣 You, via Claude<br/>(MCP server + skill)"] --> T
    B["🖥 Web console<br/>(no terminal needed)"] --> T
    C["⏰ Scheduled briefings<br/>(nobody asks)"] --> T
    T["groundstation tools"] --> D["STAC catalogs<br/>Earth Search · NASA VEDA<br/>Planetary Computer"]
    T --> E["TiTiler<br/>tiles · previews · band math"]
    T --> F["EONET · GDACS · Open-Meteo<br/>events & weather"]
    T --> G["🗺 shareable map artifacts<br/>📄 decision-ready briefs"]

Everything runs against public, keyless endpoints — it demos anywhere, with nothing to sign up for.

Real examples (all real runs, real data)

"What's burning near Chelan County?" — one scan found the two active wildfires (Navarre Coulee, Chelan Hills), correlated them with 14 straight days of zero rain and a 33°C heat peak in the forecast, and pointed at the three sub-1%-cloud scenes from ignition day.

"How did vegetation change around Wenatchee?" — one compare_dates call matched two scenes from the same Sentinel-2 tile (10TFT, both ~1% cloud) and answered: NDVI 0.46 → 0.63 between May 28 and July 5, a +36% spring green-up, with a swipe map to see it.

"Watch these four places every morning." — the fleet sweep triaged Chelan ACT (fires + heat), Efate and Laredo WATCH, Barotse CALM ("NDVI 0.397 → 0.395, normal dry-season drift, not a distress signal"). On its second run the Chelan brief opened with "unchanged since this morning's earlier run" — it remembers yesterday and only makes noise about what's new.

Related MCP server: SkyFi MCP Server

Quickstart

As a Claude Code plugin (easiest — brings the MCP server and the earth-data skill together; requires uv):

/plugin marketplace add dannybauman/groundstation
/plugin install groundstation@groundstation

First use after install takes a few seconds: uv builds the server's virtualenv on the first launch, so the tools appear a moment after Claude starts. If they don't show, run /mcp to confirm groundstation is connected — if it isn't, /reload-plugins or restart the session, and make sure uv is on your PATH.

Or skip the guessing and run the preflight:

scripts/doctor.sh

It walks the first-run chain in the order it actually breaks (uv, server env, CLI, plugin wiring, endpoints) and prints the exact fix for the first broken link. Setting up on a new machine or prepping a demo room? Run it before opening Claude.

As an MCP server directly:

git clone https://github.com/dannybauman/groundstation && cd groundstation && uv sync
claude mcp add groundstation -- uv --directory "$PWD" run groundstation

The web console:

uv run --group web groundstation-web   # open http://127.0.0.1:8765

A brief, or the morning sweep:

uv run briefing/brief.py --place "Chelan County, Washington" --days 10
uv run briefing/brief.py --fleet briefing/fleet.json                  # + --slack-webhook <url> to deliver

Updating

New tools don't reach an existing install on their own. Plugin installs are cached per version, so an installed copy stays exactly as it was until the plugin version changes and you update it.

claude plugin marketplace update groundstation
claude plugin update groundstation@groundstation

Then restart Claude Code. If you installed with claude mcp add instead, git pull in your clone and restart.

Not sure which copy you're running? scripts/doctor.sh says so, and prints the update command when the copy Claude runs is behind this one.

Things to ask it

• "Find the clearest Sentinel-2 scene of Lake Chelan from the past two weeks, tell me what's burning nearby, and give me a map I can share."

• "How much surface water is on the Barotse Floodplain right now vs early March? Use NDWI, give me numbers and a swipe map."

• "What does NASA VEDA have on the Caldor fire? Put the burn severity layer over a current scene."

• "Show the newest NAIP aerial imagery of Des Plaines, Illinois with ESA WorldCover land cover as a toggle layer."

• "Any active flood alerts along the Rio Grande between El Paso and Laredo? Alerts, week-ahead rain, latest usable imagery, one map."

• "Compare vegetation in the Yirgacheffe coffee region between January and now, and tell me which scenes you'd trust."

• "Brief me on Efate, Vanuatu: open alerts, the week ahead, the most recent cloud-free scene, all on a map."

• "Give me a 3D fly-through of Torres del Paine."

• "Make me a postcard of that I could post."

First search_datasets call takes ~20–30s while collection lists cache; everything after is instant.

What an agent gets

The paired skill in skills/earth-data/ carries the judgment layer: which catalog for what, asset conventions, index-layer recipes, "always end spatial answers with a map."

Earth briefs you

The briefing engine inverts the interaction — instead of you asking the right question, Earth reports in: fresh scenes, events, weather, an NDVI change signal against last month, a CALM/WATCH/ACT alert level, and suggested next steps, as a shareable HTML page. It keeps per-AOI memory so "what changed" means changed since the last run, fleet mode writes a triaged morning-sweep index, and --slack-webhook delivers the summary where people already look. Cron it and it runs while nobody's watching.

Scheduled sweeps

briefing/run.sh is the unattended entrypoint (unix only): it takes a lock so overlapping runs skip cleanly, appends everything to briefing/state/run.log, and runs the fleet. Every brief is checked by evals/brief_checks.py before posting — failing briefs are withheld from the Slack message and the post says "N of M areas, K withheld by checks" rather than pretending. A CALM day after a WATCH day says so in the brief, guaranteed, not just prompted.

Dry-run first, then schedule:

SLACK_WEBHOOK_URL=https://hooks.slack.com/... bash briefing/run.sh --slack-dry-run

cron (every morning at 7):

0 7 * * * SLACK_WEBHOOK_URL=https://hooks.slack.com/... /path/to/groundstation/briefing/run.sh

launchd (macOS): save as ~/Library/LaunchAgents/org.groundstation.sweep.plist, then launchctl load it.

<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE plist SYSTEM "file://localhost/System/Library/DTDs/PropertyList.dtd">
<plist version="1.0"><dict>
  <key>Label</key><string>org.groundstation.sweep</string>
  <key>ProgramArguments</key><array>
    <string>/bin/bash</string><string>/path/to/groundstation/briefing/run.sh</string>
  </array>
  <key>EnvironmentVariables</key><dict>
    <key>SLACK_WEBHOOK_URL</key><string>https://hooks.slack.com/...</string>
  </dict>
  <key>StartCalendarInterval</key><dict><key>Hour</key><integer>7</integer><key>Minute</key><integer>0</integer></dict>
</dict></plist>

Local models

Brief synthesis can run on a local OpenAI-compatible endpoint (Ollama, LM Studio) instead of claude -p:

GROUNDSTATION_LLM=auto \
GROUNDSTATION_LOCAL_URL=http://localhost:11434/v1 \
GROUNDSTATION_LOCAL_MODEL=qwen3.5-9b \
uv run briefing/brief.py --place "Chelan County, Washington"

GROUNDSTATION_LLM is claude (default), local, or auto. The local path preflights the endpoint and the prompt size, and on any failure falls through to claude -p, then to the deterministic data-only brief — the chain never produces nothing. Synthesis is the only task that goes local: it's bounded completion work, which small models do well. The agent loop's tool-calling never will — small models tend to describe tool calls instead of making them. Reasoning and boundaries: docs/adr-local-models.md.

Be a good neighbor: titiler.xyz

All tiling, previews, and pixel math ride titiler.xyz by default — a free, shared community endpoint that Development Seed runs as a demo. It rate-limits (HTTP 429) under heavy use, and a day of agent runs or a room full of people scanning places can hit that. Built-in mitigations: map artifacts carry scene footprint bounds so browsers never request out-of-footprint tiles, and statistics use small max_size reads.

When to use your own tiler — fleet briefings on a schedule, field-test-style batch runs, live demos to an audience, anything sustained:

docker compose up -d titiler                          # TiTiler is DevSeed OSS — one container (see compose.yml)
export GROUNDSTATION_TITILER=http://localhost:8000    # everything routes there

Any TiTiler deployment with the /stac router works (a hosted one, eoAPI's raster service, your own cloud instance). titiler.xyz is for kicking the tires; your own endpoint is for real work.

One caveat we learned in the field: some Earth Search collections (NAIP, Landsat) sit in requester-pays buckets. titiler.xyz carries AWS credentials for those; a local TiTiler returns 500s on them unless you provide your own AWS credentials (see compose.yml) — or just use Planetary Computer's copies of those datasets.

Reliability

Three layers of checks, because a generative product without evals is a demo, not a tool:

• uv run evals/unit_checks.py — offline, deterministic; runs in CI on every push

• uv run evals/run_evals.py — 10 live checks against the real endpoints (on-demand + weekly CI)

• uv run evals/brief_checks.py — grounding checks on generated briefs: required sections, alert level, dates cited, and every claimed event must exist in the input data (hallucination guard)

More

• docs/examples.md — the field test: twenty example prompts actually run through the agent, graded, with the improvements each round produced; visual walkthrough at /docs/field-test.html when the console is running

• docs/architecture.md — how it all fits together, design decisions, and what's deliberately not built

• Deep exploration hands off to stac-map (Pete Gadomski), which renders COGs client-side via deck.gl-raster (Kyle Barron); tiling and pixel math ride TiTiler

• A Development Seed labs prototype, July 2026