rihla

Rihla

Rihla (رحلة), "the journey." Named for Ibn Battuta's 14th-century travelogue — the record of history's farthest-traveling explorer, who left Tangier for Mecca and kept going for 29 years and 75,000 miles.

Flexible multi-leg, multi-airport flight search that finds the cheapest route across an entire itinerary — available as a CLI and as an MCP server for AI agents.

Consumer flight sites price one leg at a time. Rihla treats the whole trip as a single optimization problem:

"Leave Montevideo — or Buenos Aires, it's a ferry ride away — around September 15 for Europe. Stay 20–30 days, then Japan for 15, then home. What's the cheapest combination?"

That query has three legs, flexible airports on both ends of each, a flexible departure window, and stay-duration constraints linking the legs. Rihla prices each leg's date×airport grid once, then finds the cheapest valid combinations in pure local compute — so even a 3–4 leg flexible query costs only a handful of API calls.

1. 1,297 USD   (44 days door-to-door)
     MVD -> AMS  2026-09-08  $    383  [mock, cached]  AA AA127
     FRA -> NRT  2026-10-07  $    481  [mock, cached]  KL KL411
     HND -> EZE  2026-10-22  $    433  [mock, cached]  AZ AZ933

2. 1,308 USD   (44 days door-to-door)
     MVD -> AMS  2026-09-14  $    446  [mock, cached]  AA AA127
     ...

Note the airport substitution at work: it enters Europe through Amsterdam, leaves from Frankfurt, and flies home into Buenos Aires — each leg independently picks the cheapest airport pair from its sets.

Quick start

Requires Python ≥ 3.10.

pip install rihla            # CLI
pip install "rihla[mcp]"     # CLI + MCP server

Runs offline out of the box (no keys needed — a deterministic mock data source):

RIHLA_PROFILE=mock rihla                # the canonical demo trip above
rihla examples/queries/canonical.json   # same trip, from a query file
rihla -i                                # build a query interactively

For live prices, copy .env.example to .env and add at least a Travelpayouts token (free). See Data sources.

Writing a query

A query is a small JSON file: your origin airports, the ordered stops, a departure window for the first leg, and how long to stay at each stop. Downstream date windows are derived — you never hand-compute "if I leave Sep 15 and stay 20–30 days, when do I fly to Tokyo?"

{
  "origins": ["MVD", "EZE", "AEP"],
  "stops": ["EUROPE", ["NRT", "HND"], "MVD_AREA"],
  "earliest": "2026-09-08",
  "latest": "2026-09-22",
  "stays": [[20, 30], [15, 15]],
  "date_step": 3,
  "top": 5
}

• origins — IATA codes tried as one origin set; the cheapest wins per date.
• stops — each stop is a region name or a list of IATA codes. Built-in regions: EUROPE (MAD, BCN, LIS, CDG, FCO, AMS, FRA), TOKYO (NRT, HND), MVD_AREA (MVD, EZE, AEP). Make the last stop your origin set to fly home.
• earliest / latest — the departure window for the first leg only.
• stays — [min, max] nights at each intermediate stop (len(stays) == len(stops) - 1).
• date_step — sample every N days across date windows (coarser = fewer API calls).
• top — how many ranked combinations to return.
• currency — optional ISO 4217 code (default USD).

rihla -i walks you through these questions and prints the resulting JSON to save for reuse. Add --links to any run to show booking URLs.

Reading results

Rihla is honest about data quality rather than pretending everything is bookable:

• Each flight is tagged with its source, and cached when the price is indicative (Travelpayouts data is aggregated search history, not a live fare) versus a real, bookable fare (SerpApi / Google Flights).
• If some legs can't be priced (thin routes are real — cached sources have gaps), you get a partial result over the legs that were found, never a silently wrong total.
• For unpriced legs, Rihla shows the nearest cached fares outside your departure window — a hint to shift or widen dates.

Ranking is price-only in v0.1. Open-jaw within a region is allowed by default (enter Europe at one city, leave from another); the cost of repositioning inside the region is not modeled.

Data sources & profiles

Source	Role	Cost	Notes
Travelpayouts / Aviasales	Primary	Free	Cached, redistribution-licensed. A month of prices per call, so the call budget stays tiny. Prices indicative; coverage follows route popularity.
SerpApi (Google Flights)	Fill	250 free searches/mo (BYO key)	Real bookable fares; fills routes Travelpayouts misses. Only spent on uncovered routes.
Mock	Offline	—	Deterministic fake prices for development and demos.

Configure via .env (see .env.example) or environment variables: TRAVELPAYOUTS_TOKEN, SERPAPI_KEY, and RIHLA_PROFILE:

• local (default) — every source whose key is set, including SerpApi.
• hosted — redistribution-licensed sources only (SerpApi disabled: it scrapes Google, so do not serve it from a public hosted instance).
• mock — force the offline fetcher, no network.

With no keys set, Rihla falls back to the mock source and says so.

MCP server

Rihla ships an MCP server (stdio) so agents like Claude can run trip searches:

pip install "rihla[mcp]"
claude mcp add rihla -- rihla-mcp        # Claude Code

or in any MCP client config:

{
  "mcpServers": {
    "rihla": {
      "command": "rihla-mcp",
      "env": {
        "TRAVELPAYOUTS_TOKEN": "your-token",
        "SERPAPI_KEY": "your-key"
      }
    }
  }
}

(The server also loads a .env from its working directory, so env is optional if you run it from a checkout.)

It exposes two read-only tools with an enforced etiquette:

1. resolve_airports — the agent proposes IATA codes for the traveler's named places; Rihla validates and enriches them (nearby alternatives included). Cheap and quota-free.
2. search_trip — the priced, quota-limited search. Tool descriptions instruct the agent to get the traveler's explicit confirmation of the airports before spending quota here.

mcp-name: io.github.leojg/rihla

How it works

Leg prices are independent — the MVD→Europe fare doesn't depend on the Tokyo dates. So Rihla fetches each leg's date×airport grid once, then enumerates valid date combinations (respecting the stay constraints) entirely in memory. The combinatorial explosion lives in local compute, not in API calls.

core.py            data model + the pure optimizer (no I/O)
fetchers/          PriceFetcher protocol: Mock / Travelpayouts / SerpApi + merge
places.py          airport sets / regions
api.py             search_trip: serializable query in, result dict out
cli.py             thin CLI over search_trip
mcp_server.py      thin MCP wrapper over the same seam

Adding a data source (Duffel, Kiwi, …) is one more class implementing a one-method protocol: quote(origin, dest, day) -> Quote.

Scope (v0.1): flight search only — no lodging, no booking or payments, single adult, one cabin. Search returns booking links, never handles the transaction.

Development

git clone https://github.com/leojg/rihla && cd rihla
pip install -e ".[dev,mcp]"
pytest              # offline, no keys needed
ruff check .

License

Licensed under the Apache License 2.0 — see LICENSE.