Common Ground: movie night for two

Shared movie picks for two people, grounded in Qloo and callable by agents through MCP.

Comment

Inspiration

Choosing a movie together can become a list of one person's favorites followed by the other person's objections. Common Ground explores a narrower question: which new choices rank reasonably well for both people when their preferences differ? AI agents proposed and built this project under the participant's authorization. This is a product use case, not a personal usage story or a claim about the participant's creative development.

What It Does

Each person selects one to three films. Live title search shows years to distinguish movies with the same name. Qloo generates candidates for each person, then ranks the same combined pool separately for both. The tool returns up to three choices with each person's relative rank. The compromise rule minimizes the worse rank, then uses mean rank and entity ID to break ties. Preference films are excluded. Missing evidence for either person excludes a candidate; it is not treated as dislike. Ranks are pool positions, not enjoyment probabilities.

Why Qloo Is Essential

Live results use Qloo's taste data rather than movies or preferences invented by an LLM. Search resolves titles, and four Insights requests generate and evaluate the common pool. Affinity scores from separate queries are not averaged. An agent can call resolve_movie, confirm an ambiguous title with the user, then call find_common_ground. Both tools have been exercised through an authenticated HTTP MCP connection with real Qloo data. There is no controlled claim of better recommendation quality than an LLM-only system.

How It Was Built

React and TypeScript provide the browser walkthrough. A Cloudflare Worker calls Qloo from the server and exposes HTTP MCP. The event key remains a server secret. Preferences and results stay in browser memory and clear on reload. There is no application database, film footage, poster collection or saved audio. The source includes an MIT license, setup instructions and tests. Public browser access is an explicit server setting. Shared demo requests have a best-effort per-Worker budget; Qloo limits also apply. MCP tool calls retain sign-in requirements.

Challenges And Verification

Real API checks confirmed Search and Insights wrappers, descending numeric affinities and matching fixed-pool membership. A deployed edge-runtime failure with redirect:error was fixed by manual redirect handling that rejects unexpected redirects. Real title searches and the complete recommendation workflow passed. The deployed MCP tools resolved titles by year and returned three compromise choices. Nine checks cover ranking, missing evidence, input/response validation, credential-safe redirect rejection, MCP behavior and public-demo access/budget handling. Type checking and the production build are checked before publication. Earlier browser checks covered sample interactions on desktop and a 390-pixel layout. Sample rankings are authored illustrative fixtures, not live Qloo evidence. No user study or long-term recommendation-quality evaluation has been performed.

AI disclosure

AI agents proposed most of the product design and produced the implementation, tests and submission draft under the account owner's authorization. The participant supplied profile details, completed account verification and approved actions. The entry does not claim that the participant personally supplied most product ideas, wrote the code or performed the tests. The organizer has not confirmed whether this degree of agent involvement meets its creativity requirement.

Limitations And Next Steps

The app does not check streaming availability or guarantee either person's enjoyment. The event key has event-specific access and quota limits. Shared live access may pause during busy periods. A future evaluation with consenting pairs would test usefulness; that study has not happened.

Testing instructions

  1. Select Live Qloo in the demo.
  2. For Person 1, search Arrival and choose the 2016 movie.
  3. For Person 2, search Before Sunrise and choose the 1995 movie.
  4. Select Find shared picks and inspect the cards and individual ranks.
  5. For agent use, connect the site's /mcp endpoint, complete ChatGPT sign-in, resolve each title, confirm the intended year, then call find_common_ground with mode live and the selected IDs.
  6. Sample mode is clearly labelled and does not count as real Qloo output.

Built With

Share this project:

Updates

Submission history