Umbra Py MCP Server

umbra-py

Search, preview, load, and convert Umbra open SAR data.

Umbra publishes 16–25 cm SAR as CC BY 4.0 open data, but no search API — only a 17+ TB S3 bucket and a static STAC tree. umbra-py is that layer: search, preview, download, and analysis-ready arrays without the usual 500 lines of glue. A community STAC API (umbra serve) and MCP server sit on the same host, so pystac-client and Claude can query the archive with nothing installed.

📖 Docs: umbra-py.space · Showcase: browse the archive in the browser (no install)

Status: v0.1.2. Discovery, download, xarray loading, SICD → geocoded COG, change/timescan composites, chips, a STAC API (umbra serve, with a community host), and an MCP server all ship. This is not an InSAR toolbox (phase is not preserved through convert). Not affiliated with Umbra Lab, Inc.

Install

pip install umbra-py              # core: search + download + metadata
pip install "umbra-py[load]"      # + xarray / rasterio
pip install "umbra-py[viz]"       # + quicklooks, maps, galleries
pip install "umbra-py[convert]"   # + SICD → geocoded COG
pip install "umbra-py[all]"       # convert + load + viz + export

Python 3.10+. Other extras (dask, serve, mcp, ai, langchain, llamaindex) are listed in the install guide.

Five minutes to a scene

Fetch the weekly catalog snapshot, then search and preview offline. A live walk of the bucket (umbra search without --local) works but is slow.

pip install "umbra-py[viz,load]"
umbra index fetch
umbra search --local --area Centerfield --product GEC --limit 3
umbra gallery --local --area Centerfield --limit 6 --out gallery.html --db

from umbra_py import CatalogIndex, to_xarray

with CatalogIndex.from_release() as index:
    item = next(iter(index.search(area="Centerfield", product_types=["GEC"], limit=1)))

# Stream a downsampled window over HTTP — no multi-GB download. Needs [load].
da = to_xarray(item, max_size=1024, db=True)
print(item.summary())

If the snapshot is missing, the same search against the live bucket is UmbraCatalog().search(...) / umbra search --area Centerfield.

What you can do

More detail, options, and caveats live in the docs.

Search by bbox, place name, polygon, or Umbra task (area=). --local reads the snapshot; omit it to walk S3.

from umbra_py import UmbraCatalog

for item in UmbraCatalog().search(area="Centerfield", product_types=["GEC"], limit=5):
    print(item.summary())

Preview without downloading the scene: umbra gallery, umbra quicklook <stac-url> --out scene.png --db, umbra view <stac-url> (full-res tiles), or umbra change --area Centerfield --out change.png.

Load a geocoded GEC into xarray or a GeoTIFF (to_xarray, to_geotiff, to_stack). Needs [load].

Convert a SICD to a north-up amplitude COG (sicd_to_geocoded_cog, umbra convert) — phase is discarded. Needs [convert]. Open products generally have no radiometric metadata, so --calibrate / --noise-model measured refuse rather than invent numbers. See limitations and the complex-product handoff.

Chip scenes into georeferenced ML tiles for SR / ATR-style benchmarks from open Umbra GEC/SICD: umbra chips --area Centerfield --out chips/. See the ISR training-set cookbook and Used in research.

Drive it from an agent. Copy-paste recipes for Claude Desktop and Claude Code: Connect Claude (MCP).

Zero-install remote MCP (no uvx):

# Claude Code
claude mcp add --transport http umbra https://api.umbra-py.space/mcp --scope user

{
  "mcpServers": {
    "umbra": {
      "url": "https://api.umbra-py.space/mcp"
    }
  }
}

Paste that JSON into Claude Desktop (claude_desktop_config.json). Claude Code needs "type": "http" on the same URL — see the MCP page.

Local stdio (server on your machine):

uvx --from 'umbra-py[mcp]' umbra-mcp

{
  "mcpServers": {
    "umbra": {
      "command": "uvx",
      "args": ["--from", "umbra-py[mcp]", "umbra-mcp"]
    }
  }
}

That command is published to the MCP registry as io.github.reesehammer/umbra-mcp. STAC for pystac-client / QGIS is https://api.umbra-py.space/ (not /mcp). docker compose -f deploy/docker-compose.yml up is the one-command self-host.

What the data looks like

umbra-py downloads SICD/CPHD. umbra convert geocodes a SICD to amplitude and discards phase. It does not form interferograms or compute coherence. For a processor that needs the complex pixels, see Complex products (SICD/CPHD).

Data license & attribution

Umbra's imagery is CC BY 4.0. If you use or redistribute the data or derived products you must attribute Umbra, e.g.:

Contains Umbra open data, licensed under CC BY 4.0.

umbra-py itself is Apache 2.0 (LICENSE). The two licenses are independent and compatible.

Citing umbra-py

Machine-readable metadata lives in CITATION.cff. GitHub renders it as a "Cite this repository" button. Please also honor the CC BY 4.0 line above for any Umbra data you use.

Repo layout

Self-host: docker compose -f deploy/docker-compose.yml up (build context stays the repo root). More in docs/README.md and the deploy guide.

Community

• Contributing · Code of Conduct · Security
• Example notebooks · Limitations

Acknowledgements

Built on the SAR open-source community, including sarpy and Umbra's open data program. Not affiliated with or endorsed by Umbra Lab, Inc.