Atlas Globe

Atlas is a static spatial data workbench designed for a person and an agent to use together. It runs DuckDB-Wasm and Globe.gl entirely in the browser, keeps every dataset in one local SQL catalog, and exposes the same query and visualization actions through WebMCP.

Explore 362 IKEA locations worldwide, with each point opening a compact detail card.

The implementation deliberately stays small: vanilla HTML, CSS, and JavaScript; no application framework; no server-side database; and no runtime CDN requests. The checked-in datasets are compact snapshots with open redistribution terms and source-level provenance.

Run locally

Prerequisites: Node.js 20 or newer and npm.

npm ci
npm run serve

Open http://127.0.0.1:4173/. The local server adds the Permissions Policy and origin isolation headers required by WebMCP. The preserve lifecycle step copies pinned DuckDB-Wasm and Globe.gl files from node_modules into vendor/ before serving.

The app remains usable when WebMCP is unavailable. For agent tool testing, use a browser with the current WebMCP implementation enabled and inspect its registered tools in DevTools.

Build a static deployment

npm ci
npm run build

The generated out/ directory is self-contained and can be uploaded to Cloudflare Pages or any equivalent static host. It includes:

• all runtime JavaScript and WebAssembly;
• ten compact datasets and ten precomputed opening scenes;
• local globe and catalog imagery;
• deployment headers, licenses, and provenance metadata; and
• version.json with a b3- BLAKE3 content fingerprint.

Add ?scene=theme_parks, ?scene=earthquakes, or another scene ID to make the randomized opening scene reproducible.

What is included

The local DuckDB catalog contains ten relations:

• NOAA ocean drifter loops and Atlantic cyclone tracks;
• Natural Earth country boundaries and major rivers;
• USGS earthquakes;
• NOAA historical tsunami sources and runups;
• NOAA/SPC United States tornado tracks;
• Overture Maps locations for eight global chains; and
• Wikidata theme parks and heritage landmarks.

data/sources.json records the publisher, source and download URL, retrieval date, transformation, caveats, row count, and SHA-256 digest for every relation. See DATA_LICENSES.md for the redistribution terms and required attribution.

The repository intentionally omits raw acquisition caches, duplicated CSV exports, generated build output, and datasets without sufficiently clear open redistribution terms.

WebMCP surface

Atlas registers six imperative tools with document.modelContext.registerTool():

1. inspect_workspace
2. inspect_relation
3. load_dataset
4. execute_sql
5. render_globe_scene
6. present_result

The intended loop is:

inspect workspace → inspect relation → execute SQL → render globe scene → present result

Tools call the same domain functions as the human UI. SQL is restricted to one read-only SELECT or WITH statement, file and network readers are blocked, and saved results are capped at 25,000 rows. Scene specifications are declarative: a scene references saved result IDs and maps columns to points, paths, rings, or GeoJSON polygons without executing agent-authored JavaScript.

load_dataset can add a public CORS-enabled CSV/Parquet URL or at most 500 inline records. Imported data is labeled separately from the bundled catalog and retains the supplied provenance.

Project layout

app.js, webmcp.js       app state, DuckDB, globe scenes, and WebMCP tools
index.html, styles.css  compact responsive interface
data/                   open, browser-ready dataset snapshots and metadata
assets/                 project artwork and the local globe texture
scripts/                reproducible runtime, scene, static-build, and checks
third_party/licenses/   license texts for bundled runtime dependencies

Checks

npm run check
npm run build
npm run check:data

The data check opens every relation in DuckDB, validates expected coverage, verifies every catalog image, and checks the generated scene contract.

License

Application source and project artwork are available under the MIT License. Third-party code and datasets retain their own terms; see THIRD_PARTY_NOTICES.md, DATA_LICENSES.md, and data/sources.json.

The bundled datasets are illustrative snapshots, not operational decision-support products. Brand names identify records in the source data and do not imply endorsement.