Skip to main content
Glama

Quick Start

claude mcp add skyglance -- uvx --from skyglance skyglance

Then ask Claude: "What's flying over me right now?"

Needs uv. Its installation guide covers every platform. No API keys — every data source here is free.

Set your location so you can stop repeating it, and so SkyGlance can build a sighting history:

claude mcp add skyglance \
  -e SKYGLANCE_HOME_LAT=40.71 -e SKYGLANCE_HOME_LON=-74.01 \
  -- uvx --from skyglance skyglance

Install as a Claude Code plugin

Ships the server together with the sky skill, which teaches Claude how to read elevation angles and how to talk about aircraft to someone who has never spotted one.

/plugin marketplace add darshjoshi/skyglance-mcp
/plugin install skyglance@skyglance

SkyGlance Spotter: the Claude directory edition

The repository root is also a plugin, SkyGlance Spotter (skyglance-spotter), built for Anthropic's Claude plugin directory. It runs the same server from this repository's source and differs in four ways:

  • Two tools are left out. military_aircraft and privacy_blocked_aircraft aren't included, and search_aircraft has no military filter. Those features exist to follow aircraft whose operators or owners would rather not be followed.

  • History is opt-in. Sighting history is recorded only after you switch on Record sighting history in the plugin's settings.

  • Settings instead of environment variables. Home latitude and longitude are optional plugin settings.

  • History lives in the plugin's data folder. The history database is kept in a folder Claude Code preserves across plugin updates and deletes when you uninstall. The Python environment uv builds on first start sits in the plugin folder and is rebuilt after each update.

/plugin marketplace add darshjoshi/skyglance-mcp
/plugin install skyglance-spotter@skyglance

It needs uv. On first start, uv run --locked installs the exact dependency versions in uv.lock (the MCP SDK, httpx and their dependencies) from PyPI, which takes a few seconds. SkyGlance Spotter works in Claude Code and Cowork. It runs locally, so it isn't available in claude.ai chat.


Related MCP server: flight-tracker

What you can ask

"What's flying over me right now?"         → ranked by what you'd actually notice
"What's about to pass overhead?"           → but only 90 seconds ahead, honestly
"What is that plane?"                      → type, operator, route, photograph
"Where is flight UAL123?"                  → anywhere in the world
"Any A380s airborne?"                      → 44 of them, right now
"Any military aircraft near me?"           → and any emergencies, worldwide
"Have I ever seen this plane before?"      → your own history says
"What's the closest one ever?"             → your personal records
"Can I even see anything tonight?"         → cloud, visibility, daylight

Why it's shaped this way

"Over you" is an angle, not a distance. A radius filter returns an airliner at 38,000 ft that is 19 km away — sitting 32° above the horizon, a speck near the treeline — and gives it the same weight as a helicopter at 800 ft, at 0.7°, invisible behind buildings. SkyGlance computes the elevation angle, the compass bearing to look, and the slant range, so an answer tells you where to point your face.

It won't predict further than it can. Straight-line extrapolation from track and ground speed is accurate to about 0.11 km at 30 seconds. At four minutes the median miss is 6 km and 77% of aircraft are more than 2 km off, because aircraft on approach turn constantly. So countdowns stop at 90 seconds and anything beyond is reported as "inbound" with no number attached.

It knows when it can't see. Free ADS-B coverage comes from volunteers with receivers on their roofs. Dense Europe, US and East Asia get 90–95% of what a commercial feed shows. Oceans, Africa and central Asia get near zero. When results are thin, SkyGlance says so rather than implying an empty sky.

Your coordinates are rounded before they leave. Feed queries are coarsened to two decimal places (~1.1 km) so no volunteer feed operator gets a rooftop-accurate log of where you are and when you're awake. Bearing and elevation are still computed locally from your exact position, so what you're told to look at stays precise. There's a test that fails if an uncoarsened coordinate ever reaches a feed.


The 24 tools

Group

Tools

Overhead

whats_overhead · coming_overhead · nearest_aircraft

Identify

identify_aircraft · track_flight · track_history · airline_info

Airports

airport_activity · busiest_airports

Search the world

fleet_view · search_aircraft · global_stats

Spotting

military_aircraft · emergencies · find_by_type · interesting_nearby · privacy_blocked_aircraft

Context

viewing_conditions · feed_health

Your history

sighting_history · is_this_new · my_records · spotting_stats · poller_status

How this compares to Flightradar24

SkyGlance is a question-answering tool, not a map, so the comparison is about what you can find out rather than what you can look at.

SkyGlance

FR24

Live positions, type, registration, operator, photos

✅

Free

Squawk, vertical speed, aircraft age

✅

Silver

Filters by airline, type, altitude, speed

✅

Silver

Fleet view — every flight one airline has up

✅

Business

Airport activity, busiest airports

✅ observed

Business

Flight path history

✅ ~24 h

90 days – 3 years

Global stats

✅

—

Schedules, delays, gates, ETAs vs timetable

❌

Free

Ocean coverage (satellite ADS-B), MLAT, radar

❌

All tiers

Aeronautical charts, oceanic tracks, ATC boundaries

❌

Gold

The honest summary: for spotting and identification — what's above me, what is that, where has it been, how many 747s are up — SkyGlance covers what FR24 charges Silver and Business for, at $0. For travel logistics — is my flight delayed, what gate — it cannot compete, because that's licensed schedule data rather than ADS-B. And FR24's satellite network sees oceanic traffic that volunteer receivers simply cannot hear.

So: a strong replacement for the enthusiast half, not for the passenger half.

What it deliberately can't do

Free ADS-B carries positions, not schedules. So there are no delays, no ETAs against a timetable, no gates or terminals, and no way to search "flights from JFK to London today" — route lookup only runs the other way, from a callsign you already have. track_history is a rolling ~24-hour window, not an archive.

These aren't on a roadmap. They're a different data product (Cirium, OAG, FlightAware), and buying one would end the no-API-keys story. The tools refuse these questions explicitly rather than estimating, and the bundled skill tells Claude to do the same.

Sighting history

With SKYGLANCE_HOME_LAT/LON set, SkyGlance polls your sky once a minute in the background and records every pass. That's what makes the questions no live feed can answer possible: have I seen this tail before, is this type new here, what's the lowest anything has ever come over.

  • Stored locally in SQLite at ~/.skyglance/sky.db. Nothing is uploaded anywhere.

  • Only one process polls, even with several Claude sessions open — a pid-stamped lock enforces it, so volunteer feeds don't get hit N times over.

  • Turn it off with SKYGLANCE_POLL=0.

  • Honest limit: passes are inferred from snapshots a minute apart, so a fast low pass between polls isn't recorded. poller_status tells you the interval.

Configuration

Variable

Default

Purpose

SKYGLANCE_HOME_LAT / SKYGLANCE_HOME_LON

unset

Your location. Enables history recording.

SKYGLANCE_POLL

1

Background recording. 1/true/yes/on enable it; anything else disables it.

SKYGLANCE_POLL_INTERVAL

60

Seconds between polls (minimum 30).

SKYGLANCE_HOME_DIR

~/.skyglance

Where the database and lock live.

SKYGLANCE_EDITION

unset

directory gives the SkyGlance Spotter behaviour described above. The plugin sets it.


Where the data comes from

Seven free services, no API keys, $0/month. Full terms in NOTICE.md.

Layer

Services

Positions

adsb.lol · adsb.fi

Identity

adsbdb.com → hexdb.io fallback

Photos

planespotters.net

Conditions

Open-Meteo

Both position feeds are queried in parallel and merged, not failed over. Measured over Newark on 2026-08-02, when a third feed (airplanes.live) was still in the mix: 52 aircraft from the best single source, 61 from the union — 17% more. Merging buys coverage as well as redundancy. airplanes.live has since closed its API to anyone who hasn't contacted them first, so SkyGlance no longer queries it.

Each source sits behind a circuit breaker (three strikes, 30-second cooldown). These are volunteers paying for their own bandwidth, with no SLA and no obligation to anyone. If both fail, the last good snapshot is served flagged stale rather than an error — a 40-second-old aircraft beats a spinner.

Not suitable for anything operational. Dispatch, safety, ATC-adjacent — buy a commercial feed. The upstream terms say the same.

Network access

SkyGlance only makes outbound requests to fetch public aviation data. It sends each service only what the lookup needs: a location rounded to two decimal places (about 1.1 km), or an aircraft identifier, airline code or airport code. It never sends your conversation, files or any account details, and it uses no API keys.

Host

Why

api.adsb.lol, globe.adsb.lol (and adsb.lol)

Live positions near a location, worldwide lookups by type, callsign or registration, and ~24-hour flight paths

opendata.adsb.fi

Live positions near a location, merged with adsb.lol

api.adsbdb.com

Aircraft identity and flight routes

hexdb.io

Aircraft identity, when adsbdb has no record

api.planespotters.net

Aircraft photographs, with photographer credit

api.open-meteo.com

Cloud, visibility and daylight for viewing conditions

pypi.org, files.pythonhosted.org

Plugin only: uv installs the locked dependencies on first start

github.com

Plugin only, and only if no Python 3.10+ is installed: uv downloads a standalone Python

Each service has its own terms. NOTICE.md lists them and explains how SkyGlance meets them.

Two development-only files in this repository also make network requests. Neither the server nor the plugin ever runs them:

  • tools/build_airports.py downloads the public-domain OurAirports CSV from davidmegginson.github.io to regenerate src/skyglance/airport_data.py, and a maintainer runs it by hand.

  • .github/workflows/publish.yml runs on GitHub Actions when a release is published. It downloads a pinned, checksum-verified mcp-publisher from the Model Context Protocol registry's GitHub releases, then publishes to PyPI and the MCP registry. It authenticates with the short-lived OIDC token GitHub Actions issues to that workflow. SkyGlance and its plugins never read, store or send any credential.

Privacy Policy

This policy covers the SkyGlance MCP server, the skyglance plugin and the SkyGlance Spotter plugin.

Data collection. SkyGlance has no accounts, analytics, telemetry or crash reporting, and the author receives no data from it. It doesn't read Claude's memory, chat history or your files. It uses your location only when you give one in a question or set a home location.

Usage. A location is used to find aircraft near it. Before it leaves your computer it is rounded to about 1.1 km, so no data service receives your exact position. Where to look (direction and elevation) is worked out on your computer from the exact coordinates. Aircraft identifiers, airline codes and airport codes are sent only to the services that look them up.

Storage. Everything SkyGlance keeps stays on your computer, in a SQLite database: ~/.skyglance/sky.db for the server, or the plugin's data folder for SkyGlance Spotter. It holds cached aircraft and airport lookups and, only while sighting history is on, a record of aircraft that passed over your home location. Your home location itself is stored where you configured it: in Claude Code's settings for the plugin, or in your MCP configuration for the server.

Third-party sharing. Nothing is shared with the author or sold to anyone. The services listed under Network access receive the rounded location or identifier needed for each lookup, under their own privacy policies.

Data retention. Local data stays until you delete it. Delete ~/.skyglance, or uninstall SkyGlance Spotter, which removes its data folder. Nothing is retained anywhere else by SkyGlance.

Contact. contact@darshjoshi.com, or open an issue.


Development

uv venv && uv pip install -e ".[dev]"
pytest                    # geometry parity, merge, breakers, privacy, history

The geometry tests assert exact-digit parity with reference/overhead.mjs from the Mac app, which passes its own 11 known-answer cases. If the port ever drifts, those fail first.

MIT licensed. Not affiliated with any airline, airport, or air navigation service provider.

Available Tools

24 tools
airline_infoA

Look up an airline by its ICAO code — the prefix on a callsign.

JBU is JetBlue, BAW is British Airways, UAE is Emirates. Returns the airline name, its IATA code, country, and the radio callsign crews actually say.

Inputs: code (required).

airport_activityA

What is moving around an airport right now, observed rather than scheduled.

Buckets nearby aircraft into departing, arriving, on the ground, and overflying, based on position and climb or descent rate.

This is NOT a departure board. Nothing in this data knows a schedule, a gate, or which airport a flight is bound for — a jet climbing out near the field is inferred to be departing it. In a metro area with several airports close together, a neighbour's traffic can be misattributed.

Inputs: icao (required), radius_nm.

busiest_airportsA

Which major airports have the most traffic around them right now.

Counts aircraft on the ground or manoeuvring near each of ~1,150 major airports, from one shared global snapshot. Optionally filter to a two-letter country code such as US, GB or IN.

This measures observed activity, not scheduled movements, and it favours airports in well-covered regions — a busy airport with no nearby receivers will look quiet.

Inputs: limit, country.

coming_overheadA

What is about to pass overhead, within the next 90 seconds.

Deliberately refuses to predict further out. Straight-line extrapolation from track and ground speed is accurate to about 0.1 km at 30 seconds but misses by a median 6 km at 4 minutes, because aircraft on approach turn constantly. Anything further away is reported as inbound with no countdown rather than a number that would be wrong most of the time.

Inputs: lat, lon.

emergenciesA

Aircraft squawking an emergency code, worldwide.

7700 general emergency, 7600 radio failure, 7500 hijack. Usually zero — that is the normal and desirable answer, not a failure.

Inputs: no named parameters.

feed_healthA

Per-source health: latency, failures, and whether a circuit breaker is open.

These are volunteer-run services with no SLA. When one is down the others cover it, which is why three are queried rather than one.

Inputs: no named parameters.

find_by_typeB

Every airborne aircraft of one ICAO type code, worldwide or nearby.

Examples: A388 (A380), B748 (747-8), A124 (An-124), C17. Set near_me to filter.

Inputs: type_code (required), lat, lon, near_me, +1 more.

fleet_viewA

Every flight an airline currently has airborne, worldwide.

Give the ICAO airline code — the three-letter prefix on a callsign. BAW is British Airways, UAE is Emirates, DAL is Delta, JBU is JetBlue. Use airline_info if you only have the two-letter IATA code.

Reads one shared global snapshot rather than querying per aircraft, so it is cheap to ask repeatedly. Only aircraft within volunteer receiver coverage appear — a long-haul fleet will be under-counted while its aircraft are over oceans.

Inputs: airline (required), limit.

global_statsA

How much is flying right now, worldwide: totals, busiest types and airlines.

A single shared snapshot of everything volunteer receivers can hear. Note this is always an undercount of real global traffic — oceans, Africa and central Asia have little to no receiver coverage, so those aircraft are simply absent.

Inputs: no named parameters.

identify_aircraftA

Full identity for one aircraft: type, operator, route, and a photograph.

Give any one of an ICAO hex, a registration, or a callsign. Results are cached permanently, since an airframe's type and owner do not change.

Inputs: hex_id, callsign, registration.

interesting_nearbyB

Only the aircraft worth walking outside for.

Scores each aircraft as emergency, rare type, military, large-and-low, or a close low helicopter, and returns those above the threshold. Thresholds are deliberately harsh: under an approach path a generous filter fires constantly.

Inputs: lat, lon, radius_nm, minimum_score.

is_this_newC

Have you seen this airframe, or this type, over your location before?

Inputs: hex_id, type_code, registration.

military_aircraftB

Military aircraft currently airborne, worldwide or near a location.

Set near_me true to filter to within_km of your location.

Inputs: lat, lon, near_me, within_km.

my_recordsB

Your personal extremes: closest, lowest, most directly overhead, busiest day.

Inputs: no named parameters.

nearest_aircraftC

The closest aircraft right now, by slant range, whether or not they're visible.

Inputs: lat, lon, count.

poller_statusA

Whether background history recording is running, and what it has captured.

Inputs: no named parameters.

privacy_blocked_aircraftC

Aircraft using privacy programmes that hide their identity on most trackers.

LADD (Limiting Aircraft Data Displayed) is an opt-out US owners can request, and PIA assigns a rotating temporary hex address. These are visible here because adsb.lol publishes unfiltered community data. Positions are real; the identity is obscured by design, so do not expect a registration or owner.

Inputs: lat, lon, near_me, within_km.

search_aircraftA

Find airborne aircraft matching any combination of filters.

Filter by airline (ICAO code), aircraft type, altitude band, speed band, or military status — worldwide, or within a distance of a location. This is the general search: "any 747s above 40,000 feet", "Lufthansa aircraft below 10,000 feet near me", "anything doing over 500 knots".

Inputs: lat, lon, limit, airline, +8 more.

sighting_historyA

Every recorded pass of one aircraft over your location, most recent first.

Inputs: limit, hex_id, registration.

spotting_statsC

How much you've seen: totals, most common types, busiest hours.

Inputs: days.

track_flightB

Where a specific flight is right now, anywhere in the world.

Searches the whole feed rather than a circle. Returns nothing if the flight is not currently airborne or is outside receiver coverage — the two are indistinguishable.

Inputs: callsign, registration.

track_historyA

Where an aircraft has actually flown over roughly the last day.

Returns each airborne leg — when it started and ended, how long, how far, the highest altitude reached, and a downsampled path. This is what answers "where has this plane been today", "what route did it fly", and "how many legs has it done".

Not a flight archive: roughly 24 hours only, and long straight jumps in a path are gaps in volunteer receiver coverage rather than the route actually flown.

Inputs: hex_id, callsign, registration.

viewing_conditionsB

Cloud, visibility and daylight — whether it's worth looking up at all.

Inputs: lat, lon.

whats_overheadA

What is in the sky above a location right now, and where to look.

Returns aircraft ranked by what a person would actually notice: directly overhead first, then what is about to be, then the rest by how high in the sky they sit. Includes viewing conditions, because telling someone to look up at an overcast sky is useless.

Inputs: lat, lon, limit, radius_nm.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    B3.2/5.0

    Scored across 24 tools

    Disambiguation3/5

    Most tools have clearly different outputs, but several overlapping query patterns exist: search_aircraft duplicates the type/military filters exposed by find_by_type and military_aircraft, and whats_overhead/coming_overhead/interesting_nearby all describe nearby sky observations. Descriptions are detailed enough to resolve most confusion, but an agent could easily pick the wrong one.

    Naming Consistency3/5

    Names are all lowercase snake_case and descriptive, but they follow no single convention: some are verb_noun (identify_aircraft, track_flight), some are noun phrases (global_stats, emergencies), and some are questions (is_this_new, whats_overhead). The set is readable but stylistically mixed.

    Tool Count3/5

    24 tools is in the borderline-heavy range and the server covers multiple subdomains: live search, identity, airport traffic, personal spotting history, and feed health. Each tool has a use case, but the set could likely be consolidated without losing much utility.

    Completeness4/5

    The surface covers the skywatching domain well: find/search, identify, track current/history, airport activity, global stats, weather, and personal records. Minor gaps exist, such as no controls for the background poller and no arbitrary time-window history beyond about a day, but they do not block core workflows.

    Maintenance

    ActivitySlowing
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers