Can you be somewhere by 9? Transit schedules for AI agents
The question sounds simple, and for an agent it has been unanswerable honestly: a schedule is a table of departures that varies by day, agency and timezone, and every plausible-sounding shortcut invents one. supercarto 0.5.0 answers it from real data, with rules strict enough to print in public.
What shipped
Set one environment variable to any static GTFS feed - a directory of tables or a single .zip:
SUPERCARTO_GTFS=./city-gtfs.zip
That is the whole install. get_transit appears in the MCP
tool list, /transit answers over HTTP, and the CLI grows
supercarto transit --feed city-gtfs.zip --lat 36.913 --lon -116.77
--radius 3km. No feed configured? The tool is simply absent - like
traffic, a schedule is only as fresh as an operator's last import, so
serving one is an operator's decision, not a default.
The output is the same shape an agent already reads from a maplet, one example straight from a live run:
transit:
center: "36.913, -116.77"
radius: 3.0km
feed: sample-feed
note: Stop distances are straight-line. Schedule from GTFS feed for 2007-01-11 in America/Los_Angeles.
nodes:
STAGECOACH: {type: stop, name: "Stagecoach Hotel & Casino", tags: [bus], routes: ["30", "40"], first: 06:00, last: 06:58}
NADAV: {type: stop, name: "North Ave / D Ave N", tags: [bus], routes: ["40"], first: 06:14, last: 06:44}
edges:
NADAV->DADAN: {dist: 601m, dir: south, path: bus 40}
Routes and first/last departures come from the schedule for the queried day. The adjacency edges come from the geometry, so they survive even when the timetable does not.
The honesty rules, stated in full
Transit is adversarial to the plausible-sounding shortcut. Three rules in this release:
1. Three service states, never two. A running day shows today's
routes and departures. A date with nothing running keeps the stops, with
routes: [] and a note saying no service runs - "no transit"
is not "no stops". A feed with no calendar shows the whole network and
declares the service window unknown rather than labelling whole-feed
facts as today's schedule. An agent that can tell these apart will not
tell a rider that the line they are standing at is running.
2. Times past midnight stay past midnight. GTFS writes a
departure after the service day's midnight as 25:10.
Re-rendering it as 01:10 fits the eye and strands the person
waiting. supercarto therefore renders 25:10, in the feed's
own timezone, with the evaluated instant in the note.
3. Every caveat rides in the map. Straight-line distances are
labelled as straight-line; the feed's timezone and the schedule date
travel in the note; nothing is stated for a day the feed
does not cover. If the answering model reads only one line of the
output, it still cannot invent a schedule.
How it fits the market
Dedicated transit MCP servers exist - for the MTA, Toronto, Transitland feeds. They are worth having, and they answer a transit question an agent already knows it has. supercarto ships transit as one tool inside the maplet, next to the streets, terrain, weather and daylight of the same place, cut to one token budget with every omission recorded. The agent does not decide between "explore a place" and "read a timetable"; it walks one stop list.
On the realism of it all: the implementation adds both an RFC 4180 CSV
parser (quoted fields, CRLF, UTF-8 BOM) and a minimal zip reader (STORE +
DEFLATE) as in-repo code - zero new dependencies, same as every
supercarto release so far. The token budget is hard, re-measured on the
delivered output, with connectivity-preserving pruning and an
omitted: record of every drop. 412 tests passing, outputs
byte-identical to the Python port, and the feature was verified against
real data in all three service states.
Next
Both landed since this post: GTFS-Realtime TripUpdates decorate the
boards with live per-trip delays (and mark absence honestly — "no
prediction: unknown, not on time") since 0.5.8, and the certified
one-ride-line plan mixed with walking spans shipped as
get_transit_route in 0.5.7. Whole-metro feeds load inside
Node's default heap since 0.6.1.