open-signals
Public data, ready for maps and agents. Every reading says where it came from, when it was fetched, and what licence it carries.
open-signals has three parts:
-
A catalogue of public data sources (
catalog/sources.json). Each entry gives the API, coverage, cadence, the key it needs and its licence. Weather, water, ice, hazards, flights, ships, satellites, energy, emissions, filings, trade and more. It's the list to check before writing yet another fetcher. -
Fetchers for the sources we read today. Each one turns an upstream answer into one shape, the reading:
{ "id": "noaa-coops:9063085:2026-10-10", "at": "2026-10-10", "metric": "water_level", "value": 2.762, "unit": "ft above LWD", "summary": "Lake Erie at Toledo 2026-10-10: mean 2.76 ft above low water datum (…)", "place": { "id": "coops-toledo", "name": "Lake Erie at Toledo", "lat": 41.6936, "lng": -83.4723 }, "fields": { "station": "9063085", "datum": "LWD", "readings": 155 }, "source": { "id": "noaa-coops", "name": "NOAA CO-OPS water levels", "url": "https://tidesandcurrents.noaa.gov/…", "fetched_at": "2026-10-10T19:12:03Z", "license": "public_domain" } } -
Ways in: a CLI, a REST API with an OpenAPI description, an MCP server for agents, and GeoJSON map layers. There are no runtime dependencies; it needs Node 20 or later.
Use it
A public instance runs at https://open-signals.civops.io (REST under /v1, MCP at /mcp). It
serves only public-domain and open sources, holds no upstream keys, and is rate-limited per IP.
Try https://open-signals.civops.io/v1/layers/volcanoes?elevated=true.
git clone https://github.com/Civops-Platform/open-signals && cd open-signals # Node 20+, no npm install needed
alias open-signals='node bin/open-signals.mjs'
open-signals sources fires # search the catalogue
open-signals sources --topic water --fetchable
open-signals fetch usgs-earthquakes --min 4.5 --window week
open-signals fetch noaa-coops --coops 8518750 # any CO-OPS station
open-signals fetch open-meteo --lat 41.5 --lng -81.7
open-signals layer water_level # GeoJSON
open-signals serve --port 3290 # REST + MCP
fetch and layer run on your machine with your own keys (NASS_API_KEY, EBIRD_API_KEY). The
CLI prints each source's licence so you know what you may do with what you fetched.
REST
CESMII i3X
The same layers are served as a read-only CESMII i3X 1.0 endpoint at
https://open-signals.civops.io/i3x/v1. The address space has a root, one folder per layer, and one
object per place or event. Each object's value is its latest reading (value, quality, timestamp) with
source and licence, and its history is every reading held for that place. Subscriptions support sync
and server-sent-event streams.
CESMII's conformance suite (build 6, run 10 October 2026) rates it i3X 1.0 Compatible: 50 passed, 0 failed, 10 skipped (report). The skipped tests are the write tests; full compliance needs writes, and a server that republishes public data has nothing to write.
The spec requires authentication on data endpoints, so there is a free public read-only token:
os-public-read (send Authorization: Bearer os-public-read). GET /i3x/v1/info needs none.
Static datasets
Data that changes slowly is served as GeoJSON files rather than fetched live: wildfire burn
perimeters for the last 20 years (NIFC WFIGS and USFS MTBS, one file per state), public-land share
by county (USGS PAD-US 4.1), federal campgrounds (RIDB) and fire-weather stations (RAWS), all public
domain. GET /v1/datasets lists them with source, licence and parts; GET /v1/datasets/{id}/{part}
returns one part (a two-letter state, or all). A deployment points OPEN_SIGNALS_DATA_DIR at a
folder holding the files and an index.json; the format is described in src/datasets.js.
MCP
POST /mcp uses the streamable HTTP transport (JSON-RPC 2.0). It is stateless and needs no auth.
Every tool is read-only:
search_sources,get_source,list_hubs,list_placesget_readings,list_layers,get_layerlist_datasets,get_dataset_part
To add it to a client:
{ "mcpServers": { "open-signals": { "type": "http", "url": "https://open-signals.civops.io/mcp" } } }
Licences: what a server will republish
The code is Apache-2.0. The data is not ours to relicense. Every source keeps its own terms, and the licence travels on every reading. A running server republishes data, so it applies a gate:
A refused source answers 451 with the licence and the reason. Fetching for yourself with the CLI
is not republishing, so the gate doesn't apply there.
Fetchers today
The rest of the catalogue is listed but not fetched yet. Adding a fetcher is the most useful contribution; see CONTRIBUTING.md.
Hubs
catalog/hubs.json lists the projects we watch so we don't search from scratch each time:
- God's Eye View (Bilawal Sidhu and Sameh Khamis)
- World Monitor
- OSIRIS and WorldWideView
- the official MCP Registry and awesome-mcp-servers
- Source Cooperative, the AWS Open Data registry, Data.gov and public-apis
When a source is found through one of them, its entry records that in found_via.
Run it
npm test # offline, upstreams stubbed
npm run test:live # every fetcher against its real upstream, then REST and MCP
PORT=3290 npm start
Responses are cached in memory per source, place and parameters (2 minutes for earthquakes, up to 6 hours for weekly sources), so upstreams see polite traffic.
No warranty
open-signals passes on what public agencies and projects publish, as they publish it. Readings can be late, preliminary, revised or wrong upstream. It is provided as is (see LICENSE), and it is not for navigation, safety or any decision that cannot tolerate a bad reading.