I spend a surprising amount of time talking to ChatGPT about what to watch. Sometimes I want a twisty detective story, sometimes a sitcom, sometimes the Marvel movies in release order. I’ve even had my agent interview me over several rounds to figure out why I like the shows I like, rather than just matching a genre label.
The recommendations got good. The handoff didn’t. My agent could give me a list, but I still had to search for every title in Stremio, remember what I’d already watched, and organize the ones I wanted to come back to. I wanted it to finish the job: put recommendations directly on my Stremio Home screen and, when I ask, open a title on my TV.
The recommendation belongs where I’ll watch it
Stremio Catalog MCP gives my agent a small set of tools for creating, editing, and arranging custom Stremio catalog rows. I can say “put those mysteries in a row,” then see them alongside the rest of my library, complete with posters, runtimes, descriptions, genres, and IMDb ratings. I can move a title to the top, rename a collection, or reorganize entire Home sections by chatting instead of clicking around with a remote.
The agent does the thinking. The server doesn’t run a recommendation model or secretly choose titles. It accepts the final movie or series IDs and publishes ordinary Stremio catalogs, leaving Cinemeta to provide metadata and my existing streaming add-ons to provide playback. That separation is why I can change recommendation models without touching the TV integration.
Watch history is part of the prompt, not another spreadsheet
A recommendation that I’ve already watched isn’t very useful. The watch_history tool can check up to 50 IMDb IDs in one Stremio datastore request, including which episodes I’ve seen, and it can mark a movie or a specific episode watched or unwatched. For ambiguous series history it reports uncertainty instead of making up episode progress.
An optional title-lookup tool turns names into IMDb IDs, handling alternate releases and ambiguous titles without blindly choosing the first hit. It returns year, language, genre, and verified ratings when available. That makes the conversation much more natural: I can talk in titles, while the storage and catalog APIs work in stable IDs.
Changing someone’s Home screen deserves care
Editing a catalog entry is one thing. Reordering installed add-ons and hiding Stremio’s default rows touches account-level state, and the Stremio account API replaces the entire add-on collection when saving it. A careless implementation could wipe out the streaming providers I already use.
So the Go service fetches the current collection, preserves unrelated add-ons, makes a private backup before replacing anything, serializes concurrent sync attempts with a Redis lease, and reads the result back to verify it. Custom catalog edits use revision checks and an idempotent request ID so retries don’t silently duplicate rows. If an operation spans Redis and Stremio, it reports ambiguous outcomes instead of pretending there’s a cross-service transaction.
The final handoff is a television
I also built an optional Android TV companion because I wanted to say “open that movie on the TV” and have something actually happen. The MCP sends an IMDb ID and media type through a short-lived Firebase push to one paired TV. A tiny native receiver asks Android to open Stremio’s detail page. It can target a specific episode too, but it deliberately doesn’t autoplay anything.
The response distinguishes “Google accepted the push” from “the TV received it and requested a launch.” It can’t prove a screen rendered just because a webhook succeeded. Pairing, revocation, and device secrets stay outside the model’s tool surface.
This is a personal Stremio control plane, not a streaming service. One deployment controls one account, and it doesn’t provide streams, replace add-ons, or guarantee a title is available. The part I enjoy is much less grand than the infrastructure: I ask my agent what to watch, and the result is already on the screen where I want it.
