// the take
A collection of 42 Claude Code/OpenCode/Codex skills plus 6 MCP servers for comparing cash and award flight/hotel prices across airlines, hotel portals, and loyalty programs, including data-heavy reference skills on alliances, transfer partners, and award sweet spots. It's aimed at people who already do manual points-and-miles optimization and want an agent to run the multi-source search and math for them.
The skill architecture is genuinely well thought out: tool skills wrap APIs/scrapers, reference skills carry the institutional knowledge (alliance reachability, devaluation history, stopover rules) and load only on demand via progressive disclosure, so the always-on context stays small even with 42 skills installed. It wraps Seats.aero to get award availability across 27 mileage programs in one place, which is otherwise manual, tab-heavy work. Free MCP servers (Skiplagged, Kiwi, Trivago, Ferryhopper, Airbnb) give working cash-price search with zero signup, and the install path is a single plugin command for Claude Code/Codex rather than a full clone-and-configure.
Several skills (Southwest, American Airlines, Chase, Amex, TicketsAtWork, Sutochno, VRBO) work by handing the agent your airline/bank/credit-card login credentials and driving a Patchright browser inside Docker — that's real credential exposure and ToS risk against financial accounts, not just a scraping inconvenience, and breaks the moment any of those sites change their login or booking flow. Full functionality requires stacking multiple paid subscriptions (Seats.aero Pro, AwardWallet Business, Duffel/Ignav, etc.), so the 'drop-in' pitch undersells the real cost to get award search and loyalty balances actually working. The reference data (points valuations, transfer bonuses, sweet spots) is static JSON that goes stale as programs devalue or change partners — there's a freshness-check script but the advice is only as good as the last refresh, and giving financial guidance off a few-months-old valuation table is an easy way to steer someone wrong. It's effectively a single-maintainer project with thin automated test coverage (a handful of unit tests for a couple of the fragile scraper skills), so correctness at this scale is mostly resting on manual verification.
We read the README, the directory tree and a key source file in October 2026. Nobody pays to be listed. How we judge →