Case study: RoTV Guide — teaching a model to choose among 14 tools

The second case study in the series about the four applications built for the OpenAI ecosystem. If Când Reciclăm was about honesty in data, RoTV Guide is about something else: how you teach a model to choose correctly among 14 tools, and how you keep data fresh when it expires by the hour. This is not a study about television — it's a study about AI orchestration.

At a glance

• 14 public MCP tools (in three waves of evolution) + 4 WebMCP tools running directly in the browser
• 260 TV channels, ~26,000 programs in the current schedule window, automatically refreshed several times a day
• 4 verticals in one portal: TV, streaming, cinema, theater
• Freshness is verifiable: a dedicated tool tells anyone how recent the data is
• The first app in the series submitted to the OpenAI review (July 16, 2026, in review) · MCP server in the official registry · agent-readiness score 79/100

The problem

Romanian entertainment information was scattered across dozens of sites: TV listings in one place, streaming charts in another, cinemas and theaters on their own websites. And for AI agents, all of it was completely invisible: a model asked "what's worth watching tonight on Romanian TV?" had no way to know — and would answer anyway, making it up.

Why existing solutions weren't enough

Classic TV guides are built for browsing, not for asking. They show you the grid — but they don't answer the real question, which is almost never "what's on channel X at hour Y?", but "what should I watch tonight?", "is anything important on today?", "what do we both pick?". And there was no bridge between verticals: nobody compared the movie on TV with what entered this week's streaming chart or what's playing in cinemas.

Objectives

1. One portal for TV, streaming, cinema and theater — with data that updates itself.
2. Built for the agentic web from day one: the same data accessible to humans (the website) and to AI agents (MCP, WebMCP, structured data).
3. From question to decision: not just "what's on TV now", but recommendations with criteria, comparisons and, at the limit, "you choose for me".
4. Guaranteed, verifiable freshness — in a domain where yesterday's data is already useless.

The architecture

Three layers, one source of truth:

• Collection and normalization: TV listings for hundreds of channels, the streaming platforms' official charts, cinema and theater programs — heterogeneous formats, automatically normalized into a single data model, published as a rolling window of a few days.
• The website (tv.madeinro.eu): the guide for humans — on TV now, prime time, recommendations, cinema, theater.
• The agentic layer: a public MCP server with 14 tools reads exactly the same published data as the website; additionally, 4 WebMCP tools run directly in the browser (an experimental Chrome standard), and a REST layer with API keys opens the same data to any other kind of client.

As with Când Reciclăm, the principle is unchanged: there are no different versions of the truth. The website and the agents read the same data.

Why MCP and not a classic chatbot?

For a TV guide, the answer has an extra nuance compared to the previous study. It's not only that the application goes where users already are — a conversational TV guide is, by its nature, an orchestration exercise: "what should I watch with my partner tonight, something relaxing, not too long?" requires understanding preferences, filtering the current schedule and making an argued recommendation. The model does the reasoning; the tools give it the real raw material — today's schedule, not the one from its training data. Without tools, any model will "recommend" shows that don't exist or ended long ago.

The complete flow

User
    │  "I don't know what to watch tonight. You choose."
    ▼
ChatGPT
    │  picks the right tool out of 14
    ▼
MCP server — 14 tools, 3 waves of evolution
    │  tv_concierge: criteria + current schedule → argued decision
    ▼
The guide's published data
    │  260 channels · ~26,000 programs · rolling window
    ▲
    │  automatic refresh, several times a day
Public and official sources
    (TV listings · streaming charts · cinema · theater)

The 14 tools arrived in three waves, and their evolution is the story of the product:

• Wave 1 — questions: what's on TV now, search a program, prime time, details about a title.
• Wave 2 — taste: recommend by mood, plan an evening, compare options, find something for two, explain why you recommended this, how fresh is the data.
• Wave 3 — decision: tv_concierge ("you decide for me"), what's important today, the visual card of the evening.

From "what's on TV?" to "you choose" — the road from a guide that informs to an assistant that decides.

The challenges

• Perishable data: a TV guide expires continuously — what was valid in the morning is history by evening. The solution: automatic refresh several times a day, a rolling schedule window, and a public freshness tool so any agent can tell the user how recent the data is.
• Normalization: hundreds of channels, different formats, inconsistent titles across sources. Everything is automatically brought to a single data model — that alone is what makes cross-vertical comparisons possible.
• Choosing the right tool out of 14: with this many tools, the risk is that the model picks the wrong one. The antidote was design: each tool does exactly one thing, and its name and description say precisely what — tv_now_on_tv cannot be confused with tv_plan_evening.
• Trust in recommendations: a recommendation without an explanation is an order. A dedicated tool explains the criteria behind every suggestion.
• Visual answers in chat: for the "card of the evening", the server also delivers visual interfaces (MCP-Apps), not just text — where the client supports them.

The key technical decisions

1. The same published data for everyone — the website and all agentic layers read the same data artifacts; zero divergence.
2. 14 narrow tools instead of one universal tool — tool selection is a product-design problem, not a prompting problem: the right granularity makes the model's choice almost impossible to get wrong.
3. Freshness as a public feature — not a marketing promise, but a tool anyone can call.
4. Three agentic access paths — MCP for conversational clients, WebMCP for in-browser agents, REST with API keys for classic integrations. Same source, any client.
5. Read-only, no account — the guide informs and recommends; it asks for nothing.
6. Visible incremental evolution — waves 1→2→3 shipped one at a time, each on the foundation of the previous one, without breaking anything that existed.

The result

RoTV Guide is today the first Romanian TV guide built for the agentic web: 260 channels and ~26,000 programs in the current window, plus streaming, cinema and theater, in one portal — tv.madeinro.eu. The MCP server is public, listed in the official MCP registry, with an agent-readiness score of 79/100 (12 of 14 checks). On July 16, 2026 it was the first of the four applications submitted to the OpenAI review for the ChatGPT catalog — where it currently is. The oldest application in the series (started in December 2025 as a classic TV guide) became, in July, the laboratory where the agentic architecture later used by all the others took shape.

What would I improve in the next version?

• Persistent personalization: an optional taste profile, so recommendations pick up where they left off — today every conversation starts from zero.
• Visual cards everywhere: extending visual answers beyond the "card of the evening" as MCP clients support them.
• Agent-style notifications: "the show you follow starts in 15 minutes" — the step from a reactive assistant to a proactive one.
• Finer vertical coverage: more sources for theater and cinema, where the local offer is fragmented.

---

This case study is part of a series on building AI applications for the OpenAI ecosystem. Previous episode: Când Reciclăm. Coming up: Digital Compass and Travel Trends. The context for the whole story: AI adoption is a delivery discipline.