A year ago, connecting an AI agent to live flight data meant writing your own tool wrappers. Now there are several ready-made Model Context Protocol servers, from official vendor releases to weekend projects, and they differ more than their landing pages suggest.
This is a comparison of the ones worth knowing about. One of them is ours, so we've tried hard to make the table something you'd trust even if you already knew that. Every tool count and capability below comes from each project's own README as of September 2026. These projects move quickly, so check the repositories before you commit.
The short version
| Server | Maintainer | Tools | Live positions | Schedules & status | Aviation weather | NOTAMs | Key needed | Transport |
|---|---|---|---|---|---|---|---|---|
| AirLabs MCP | AirLabs (official) | 10 | Yes | Yes | No | No | AirLabs key | stdio (npx) |
| VariFlight MCP | VariFlight (official) | 8 | By registration | Yes | 3-day airport forecast | No | VariFlight key | stdio (npx) |
| SkyGlance | Community | 24 | Yes | No | No | No | None | stdio (uvx) |
| adsb-mcp-server | Community | — | Your own feeder | No | No | No | None | Self-hosted |
| SkyLink MCP | SkyLink (us) | 46–58 by plan | Yes | Yes | METAR, TAF, SIGMET, PIREP, winds aloft | Yes | SkyLink key | Remote HTTP, or stdio bridge |
The columns that separate these servers are weather and NOTAMs. Most aviation MCP servers are flight-tracking servers. If your agent needs to answer "can I fly into Aspen this afternoon?", the position data alone won't get it there.
The servers, one by one
AirLabs MCP
AirLabs' official server is open source under the MIT licence and runs over stdio via npx @airlabs-co/airlabs-mcp. It exposes ten tools: get_flight_status, get_airport_schedule, monitor_delays, track_live_flights, find_nearest_airport, search_airport_code, get_airline_info, get_airport_info, lookup_aircraft and find_routes.
Good at: the core flight-tracking questions. Where is this flight, what's departing this airport, which airlines fly this route, what aircraft is this. Ten tools is also a small, well-named surface, which helps the model pick the right one.
Doesn't cover: aviation weather and NOTAMs. There's no METAR, TAF or airspace-notice tool in the list.
You'll need: an AirLabs API key. They offer a free signup key; check their pricing for the limits.
VariFlight MCP
VariFlight's official server is ISC-licensed and runs via npx @variflight-ai/variflight-mcp. Its eight tools lean towards travel rather than operations: searchFlightsByDepArr, searchFlightsByNumber, getFlightTransferInfo, flightHappinessIndex, getRealtimeLocationByAnum, getFutureWeatherByAirport, searchFlightItineraries and getFlightPriceByCities.
Good at: travel-planning questions. Connections, itineraries, prices, and a "happiness index" covering punctuality, cabin configuration and amenities. If you're building a travel assistant, that's a distinctive set.
Worth knowing: it's China-centric, using city codes like BJS and SHA. Live position lookup works by aircraft registration rather than by area. The weather tool is a three-day airport forecast, which isn't the same thing as a METAR or a TAF — it's the difference between a travel forecast and an aviation one.
You'll need: a VariFlight API key from their MCP portal.
SkyGlance
SkyGlance is a community project and the most charming server on this list. It's MIT-licensed, needs no API key, costs nothing, and installs with uvx. Its 24 tools are built for plane-spotting: whats_overhead, coming_overhead, identify_aircraft, military_aircraft, emergencies, privacy_blocked_aircraft, and a set that logs passes over your home location to a local SQLite database so you can ask "have I seen this tail before?"
It pulls positions from the free community feeds adsb.lol and adsb.fi, identity data from adsbdb.com with hexdb.io as a fallback, photos from planespotters.net and viewing conditions from Open-Meteo.
Good at: exactly what it says. "What's that plane over my house?" is a delightful thing to ask an agent, and it answers well.
Doesn't cover, by its own account: the README is admirably direct. It deliberately has no schedules, delays, gates, terminals, ETAs against a timetable, route search or oceanic coverage, keeps only about 24 hours of history, and states plainly that it is "not suitable for anything operational". That's the right call for a free tool built on volunteer feeds, and worth taking at its word.
Self-hosted feeder servers
If you run your own ADS-B receiver, projects like adsb-mcp-server expose your feeder's data to an agent directly. Coverage is whatever your antenna can hear, and there's no dependency on anyone else's service. It's a good fit for hobbyists who already feed a network and want to query their own data.
SkyLink MCP
Our MCP server exposes the full v3.1 API as tools. On the direct channel it's a remote, stateless streamable-HTTP endpoint at data.skylinkapi.com/mcp, authenticated with the same x-api-key header as the REST API, so there's nothing to install. On RapidAPI it runs through their MCP hub with the mcp-remote stdio bridge.
Good at: breadth. Live ADS-B, flight status, schedules, airports, airlines and aircraft, plus METAR, TAF, SIGMETs, PIREPs, winds aloft and NOTAMs, the ML block-time model, carbon estimates, and on Pro and above, historical ADS-B. It's the only server in this comparison whose tool list covers aviation weather and NOTAMs, which is what an agent needs for preflight or dispatch reasoning.
The trade-off: breadth has a cost. Depending on plan, your agent sees 46, 52 or 58 tools, and tool-selection accuracy falls as the list grows. The tool names are generated from API operations, like search_airport_airports_search_get, rather than written for a model to read. For a narrow agent, you'll usually get better results by putting a thin wrapper in front that exposes the five or six tools you actually use. Giving your AI agent real-time flight data covers how.
You'll need: a SkyLink key on a paid plan — Basic exposes 46 tools, Pro 52 and Ultra or Mega 58, with the extra tools covering historical ADS-B. The free trial (1,000 requests a month, reviewed rather than self-serve) is a way to test the underlying API before you choose a plan.
What to compare, beyond the table
Weather and NOTAMs, or just positions? This is the biggest divide. A flight-tracking server can tell an agent where a plane is. It can't tell it whether the destination is below minima or whether the runway is closed. If your agent needs to reason about flying rather than observe it, you need both.
How many tools, and how they're named. More tools isn't better. Ten clearly named tools will be used more accurately than fifty generated ones. If you choose a broad server, plan to curate its surface.
Local or remote. stdio servers run on the user's machine, which suits a desktop assistant and needs a Node or Python runtime. A remote HTTP server needs no local install, which suits hosted agents and teams, but every call goes over the network.
Whose data, and under what terms. Free community feeds are wonderful and come without an SLA, a support contact or a commercial licence. That's fine for a personal assistant and a real problem for a product. The same issues apply to MCP servers built on those feeds as to the feeds themselves — see the best free flight tracking APIs for the licensing side.
What happens to your quota. Agents don't call APIs the way apps do. One question can fan out into a dozen tool calls, and an agent stuck in a retry loop can burn through a small quota fast. Whatever server you pick, cap tool calls per turn and cache anything that changes slowly.
Which one to use
You want to ask Claude what's flying overhead: SkyGlance. Free, no key, built for exactly that.
You run your own receiver: a self-hosted feeder server, so your agent queries your own data.
You're building a travel assistant focused on China: VariFlight's pricing, itinerary and comfort tools are unusual and useful.
You need solid flight tracking and schedules and nothing more: AirLabs' ten tools are a clean, focused surface.
Your agent needs weather and NOTAMs alongside flight data — a briefing assistant, a dispatch helper, anything that reasons about whether a flight can operate rather than just where it is: that's the gap SkyLink fills in this list. Building an AI flight dispatcher works through one end to end.
Nothing stops you combining them, either. MCP clients can connect to several servers at once, and a spotting server plus an operational one is a perfectly reasonable setup. The main thing to watch is the combined tool count your model has to choose from.