Use this local MCP server to look up SanAlert's current inferred air-threat level for a place in Poland and the context needed to interpret it.
Call sanalert_level with a SanAlert town (Polish letters optional) or latitude/longitude, plus optional language pl/en/uk.
Get the level (quiet, activity, approaching, near, no_data, out_of_range), localized level name, and website sentence.
Get supporting details: reason code/kind, km/ETA/course, oblast/raion/threat, risk/confidence/degraded.
Check data trust and freshness: data, data age/timestamps, expiry, and no_data_reasons/no_data_codes; no data is never quiet.
Get official RCB/RSO messages in force for the place's voivodeship when available.
Get place resolution (town/point, grid cell, voivodeship) and required notice, attribution, and pass_through text.
Optionally access sanalert://towns for accepted town names and coordinates (per README, 1.1.0).
Privacy: the queried place is not sent; the server fetches public SanAlert state/grid documents locally.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@sanalert-mcp what's the air threat level in Przemyśl right now?"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
The unofficial threat level from the air for a place in Poland, for an AI assistant that can run a local MCP server (Claude Code, Claude Desktop and others). Ask it "what is the level in Przemyśl?" and the assistant gets the answer the website would show for that place. Not an official warning system: it does not replace sirens, Alert RCB or instructions from the authorities.
It runs on your computer and talks to the assistant over stdio.
When the assistant asks about a place, it fetches https://api.sanalert.pl/v1/state, the same
public document the website and the apps read: at most once every 5 s, with If-None-Match and
the User-Agent SanAlert-MCP/<version>. Once per run it fetches the public grid
(/v1/grid/1.json) to find the place's ~11 km square and its voivodeship.
It looks the place's level up in that document with the rules SanAlert's website and apps follow (sanalert.pl/jak-to-dziala). The place stays on your computer: no request carries it, and a town's name is never looked up online.
The privacy page says the same (sanalert.pl/prywatnosc, "Serwer SanAlert dla asystentów AI"). Terms (Regulamin, Умови).
Related MCP server: machine-mcp
sanalert_level takes place, one of the website's towns (Przemyśl, Lublin, Rzeszów, Chełm and
the rest of its list; Polish letters optional, and Warsaw and Cracow work too), or latitude and longitude, and language:
pl, en or uk.
Field | What it is |
|
|
| the level's name and the website's sentence under it, in the chosen language |
| what raised the level, as the website shows it |
| as in the state; |
| whether the level can be trusted now, and why not: |
| the town or the point, its square ( |
| RCB and RSO messages in force for the voivodeship; |
| what has to stay with the data wherever it is shown |
The towns place accepts are also a resource, sanalert://towns (1.1.0), with their coordinates.
Blind is not calm. When the document is missing, cannot be read or is older than 60 s, or the
grid cannot be read, the level is no_data, never quiet, and say says so. The age is never less
than the time the server has held the document, so a computer clock that runs behind cannot keep an
old "quiet" fresh while the API is down. A point outside
Poland's grid is out_of_range. A name that is not on the list is an error that names the list:
give coordinates instead.
For the whole of Poland (the state summarised, official messages of the last 24 hours, night
reports, the reliability record) SanAlert runs a remote MCP server, https://api.sanalert.pl/mcp.
It takes no place.
It needs uv.
Claude Code:
claude mcp add sanalert -- uvx sanalert-mcp
claude mcp add --transport http sanalert-poland https://api.sanalert.pl/mcp # optional: the whole of PolandOr both at once, as a Claude Code plugin: /plugin marketplace add plrojek/agent-skills, then
/plugin install sanalert@plrojek. Straight from GitHub instead of PyPI:
uvx --from git+https://github.com/plrojek/sanalert-mcp sanalert-mcp.
Claude Desktop (claude_desktop_config.json):
{
"mcpServers": {
"sanalert": {
"command": "uvx",
"args": ["sanalert-mcp"]
}
}
}Whoever shows or repeats the data keeps notice visible with it and a visible link to every
attribution entry, always refers readers to the official air-raid alarms (sirens, Alert RCB, the
instructions of the authorities), keeps a visible link to NEPTUN (https://neptun.in.ua, for example
"Дані: Карта повітряних тривог — NEPTUN") next to the level, never turns no data into quiet, and
says that the level is SanAlert's inference, not an official warning. In full:
terms and the API description.
uv run pytest runs the tests; none of them goes online.
Three files in src/sanalert_mcp/ are generated, never edited by hand:
core.py is the Home Assistant integration's core.py, the same rules in the same words;
towns.json is the apps' town list (ios/SanAlertKit/Sources/SanAlertKit/Resources/towns-v1.json,
made from the website's web/src/lib/cities.ts);
api_texts.json is the API's notice, attribution and pass-through (api/sanalert/state.py,
api/sanalert/mcp_server.py).
python scripts/sync.py copies the first two and python scripts/sync.py --check fails on any
difference (CI runs it). The API's tests fail when api_texts.json no longer matches the API.
The source is integrations/mcp/ in SanAlert's own repository; the public repository
(github.com/plrojek/sanalert-mcp) is a copy of it, refreshed with each release, so a fix in
core.py reaches users with the next release.
MIT, Applied AI sp. z o.o.
SanAlert's level right now for one place in Poland, worked out on this computer: quiet, activity,
approaching or near, the website's sentence about it (say), what raised it, the official RCB and RSO
messages in force for the place's voivodeship, and whether the data can be trusted now (data: ok,
delayed or no_data; no_data is never quiet). Give one of SanAlert's towns or the coordinates. The place
is never sent anywhere: the server fetches the same public documents everybody gets.
Inputs: place, language, latitude, longitude. Returns: object
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
Scored across 1 tool
Only one tool exists, so there is no risk of selecting the wrong tool. Its description clearly defines a single purpose: retrieving SanAlert level and related alerts for a specified place in Poland.
The single tool name sanalert_level is descriptive, uses snake_case, and aligns with the server name. With only one tool, there is no inconsistent convention to compare.
A single tool is thin for an MCP server, even though the tool itself is comprehensive. The rubric treats 1-2 tools as borderline, and there is no ability to split concerns across multiple operations.
The tool covers current level, messages, and data trust for one place, but there is no way to discover valid town names or coordinates, nor to query multiple places or historical data. This is a notable gap for an agent that does not already know the exact input.
Persistent memory, hybrid search and a goal graph for AI agents, over stdio or remote HTTP.
Search Bluesky posts and feeds; read profiles, threads, and trends. STDIO or Streamable HTTP.
Public AI web-readiness scanner and machine-facing observability discovery service.
US weather & geo for AI agents: forecasts, alerts, earthquakes, elevation, geocoding. No keys.
