splitfare 🎟️

A flight-hacker's scanner for short trips: split tickets, trip-shape search, and it knows what's on when you land.

Booking sites answer "how much are flights on these dates?" splitfare answers the question you actually have:

"Find me any weekend next month — out Friday or Saturday, landing before 2pm, home Sunday night — and tell me who's playing while I'm there."

It prices the direct fares and every viable split-ticket combination (separate cheap tickets glued together through hub airports — routings no booking site will show you), sweeps whole months by trip shape, and pairs every candidate window with the destination's event calendar.

Demo above: London → Berlin weekends in September for 2 people — every Fri→Sun window priced, with Berlin's club calendar (Resident Advisor) per night.

What it does

• Split-ticket routing — direct fares plus self-transfer combos through your chosen hub airports (same airport, enforced minimum connection, home the same day), ranked by total price for your whole party.

• Trip-shape search — sweep a month for the cheapest N-night window with human constraints: out on these weekdays, land by this time, home by that time.

• Overnight positioning — when landing early is impossible same-day, it prices flying to the hub the evening before and catching the morning flight.

• Multi-source fares — Google Flights plus the airlines' own public fare APIs (Ryanair, easyJet, Wizz), fetched in parallel and merged, deduped by flight, cheapest fare wins. Each source has its own health check.

• Price intelligence — every route-date keeps a price history; results show a buy/wait verdict vs its own typical range (good/fair/typical/pricey) and a rising/falling/stable trend.

• Watch mode — set a trip shape + target price ("any 24hr window in September under £180 for 2") and get a Telegram ping the moment one appears. Cron-friendly.

• Events fusion — every candidate night shows who's playing, via pluggable providers (Resident Advisor worldwide, Ticketmaster, Skiddle, Ibiza Spotlight — or write your own in ~30 lines). --events-filter big turns the calendar into a search constraint: only windows with a tier-1 night.

• A phone dashboard — publish renders a static site: browse every window by day, tap into flights, alternatives and lineups. Host it anywhere.

• AI-agent native — an MCP server exposes the whole thing to any MCP-capable assistant (Claude, etc.), plus an ask command for plain English from the terminal.

Related MCP server: Amadeus Agent

Quickstart

git clone https://github.com/teatimedev/splitfare && cd splitfare
python3 -m venv .venv && .venv/bin/pip install -e ".[mcp]"
.venv/bin/splitfare init     # your airports, hubs, party size, events provider

Then:

# price one out/back pair — direct + split combos + what's on
.venv/bin/splitfare window 2026-09-04 2026-09-06 --events

# sweep a month: cheapest 2-night windows, Friday departures, land by 2pm
.venv/bin/splitfare scan --month 2026-09 --nights 2 --out-dow fri --arrive-by 14:00

# render the phone dashboard (static site in ./site — host anywhere)
.venv/bin/splitfare publish --month 2026-09 --nights 2

# plain English (drives the tool via the `claude` CLI if you have Claude Code)
.venv/bin/splitfare ask "cheapest weekend next month for 2, landing before 2pm?"

The first scan of a month is slow (~1.5 s per route-date, politely rate-limited). Results cache for 6 h; after that everything is instant.

Picking your hubs

Hubs are what make split ticketing work: airports with cheap budget-airline service from both your home and your destination. Think Ryanair / easyJet / Wizz bases — Stansted, Luton, Barcelona, Bergamo, Vienna, Warsaw… 5–10 is the sweet spot; every hub adds ~4 lookups per date, so more hubs = more combos found but slower scans. The init wizard walks you through it.

Using it with your AI agent (MCP)

splitfare ships an MCP server, so any MCP-capable assistant can hunt flights for you:

claude mcp add splitfare -- /path/to/splitfare/.venv/bin/splitfare-mcp

Three read-only tools: splitfare_find_flights (one date pair, full detail), splitfare_scan_month (cheapest windows across a month), splitfare_whats_on (event calendar). Your agent gets structured JSON — routings, per-leg booking links, connection gaps, lineups — and can reason about trade-offs ("£15 more but you'd land 6 hours earlier and Four Tet is playing").

Three things to know:

1. Run splitfare init first — the server reads the config.json next to the code (or at $SPLITFARE_HOME). Without it, tools return a setup hint.

2. Uncached calls take seconds-to-minutes; scan_month with deep=0 on a cold cache can exceed some clients' tool timeouts — start with specific dates or deep=4.

3. Prices are totals for the configured party, in your configured currency.

Configuration

splitfare init writes config.json:

SPLITFARE_HOME relocates config/cache/site for pip installs.

Events providers

Adding one is a single function in providers.py.

Flight sources

Primary data comes from Google Flights via fast-flights — the only free source that aggregates all the budget carriers. On top of that, splitfare merges in the airlines' own public (no-key) fare APIs, configured in "flight_sources":

All enabled sources are fetched in parallel (per-source concurrency budget — polite but ~4-5x faster than the old sequential sweep) and merged, deduped by flight, keeping the cheapest fare. Fetches are cached; the cache refetches automatically when you add a source that wasn't there before.

IP-dependent sources: Ryanair works from anywhere, but easyJet/Wizz sometimes block datacenter IPs (VPSes). If a source shows as empty in the per-source health warning, run from a residential IP or drop it from flight_sources — google alone still covers those carriers via the aggregator. easyJet's current API (no public endpoint; Akamai-protected) is reverse-engineered in docs/easyjet-api.md, with EASYJET_COOKIES session-cookie injection for residential use.

For the curious: Amadeus retired its self-service API in 2026, Kiwi's Tequila closed to new signups, and Duffel doesn't carry Ryanair/Wizz — scraping the aggregator remains the pragmatic play, which is also why this is a run-it-yourself tool and not a website.

Price history, buy/wait scoring, and watches

Every fetch appends the cheapest price to .cache/price_history.jsonl (backfilled from older caches on first run). With ≥3 observations for a route-date you get a buy/wait signal on every result:

• verdict vs the route-date's own history: good (≤ p25) / fair (≤ p50) / typical / pricey (> p75)

• a rising/falling/stable trend from recent observations

The dashboard shows "vs typical" badges and a sparkline of the last 24 price observations on each flight ticket.

Watch mode turns "run a scan" into "ping me when it happens":

splitfare watch add --month 2026-09 --nights 1 --max-total 180 --name "Sep 24hr"
splitfare watch list
splitfare watch check          # cron-friendly; alerts on new/cheaper hits only
splitfare watch remove ID

watch check scans each watch's trip shape and alerts when a window totals at or below the target — only on first appearance or a ≥5% price drop, so cron doesn't spam you. Alerts go to Telegram when configured:

"alerts": { "telegram": { "bot_token_env": "TELEGRAM_BOT_TOKEN",
                           "chat_id": "123456789" } }

(create a bot with @BotFather; your chat id comes from @userinfobot). Without a token, watch check prints to stdout and exits 0 — safe for cron either way. Example cron: 0 8 * * * cd ~/projects/splitfare && .venv/bin/splitfare watch check.

Events as a constraint

scan/publish/watch take --events-filter any|big: only keep windows whose night has any event, or at least one tier-1 venue (Ushuaïa, Hï, Pacha, Amnesia, DC10, Eden…). "Cheapest 24hrs in Ibiza with a big night on" is one command now.

How it works, honestly

• Any source can break. Treat prices as estimates and book each leg directly with the airline — every result links to a matching search. The tool tracks per-source health (live vs empty fetches) and warns you per source instead of silently showing "no flights".

• Split tickets carry real risk. Separate bookings mean no missed-connection protection. The tool enforces a minimum gap and shows the gap on every result, but the risk is yours. Prefer long gaps.

• Be polite: requests are rate-limited (per source) and cached. Don't hammer it.

Development

.venv/bin/pip install -e ".[dev]"
.venv/bin/pytest tests/ -q

PRs welcome — events providers for your scene, flight sources, and routing improvements especially.

License

MIT.