Talk to film. Twenty sources, one honest score.
Choosing a film means opening four tabs and trusting none of them. Fi1m aggregates more than twenty review sources into a single blended score, writes its own reviews, and gives you a discovery interface you talk to: what you've seen, what you want to see, what you're in the mood for tonight. It's a platform for people and an MCP for their agents, so your assistant can use it too.
Three years of groundwork and a broken discovery experience
The Fi1m team had spent three years on the problem: review scores that disagree, aggregators that flatten them, recommendation engines that optimise for whatever a platform wants you to stream. Discovery had become a chore. The frontier report tracked two things arriving at once: models good enough to read and reconcile criticism at scale, and MCP as the way agents would consume services. Fi1m was the concept that used both.
A corpus from nothing, a blended score, a critic of its own, and a conversation
There was no dataset to license, so the first thing built was the corpus: twenty thousand titles, each resolved to a single canonical record across remakes, re-releases, regional titles and the twenty-plus review sources that all spell them differently. On top of it, reviews aggregated and weighted into a single blended score that shows its working, and Fi1m's own reviews alongside, so the platform has a voice rather than just an average. Then a discovery interface you talk to, which builds and maintains lists of what you've seen and what you'd like to see, and an MCP server exposing the same catalogue and lists to agents, so “find us something for tonight” works from whatever assistant you already use.
Training and indexing came first. The product sat on top.
Most of the eight months went into the corpus, because everything else depends on it. Ingestion pipelines pulled metadata, cast, crew and criticism from more than twenty sources; entity resolution collapsed them into one record per title, with every claim carrying its source. Models were then trained on that material rather than on the open web: readers fine-tuned to extract a verdict and its reasons from each critic's house style, calibration so a three-star from one source and a 7/10 from another land on the same scale, and a review model trained on editorial to write in Fi1m's own voice. The whole catalogue is indexed for meaning as well as metadata, so “something like Heat but quieter” is a query that resolves. The blending itself stays deterministic and explainable: a score can always be traced back to who said what. The MCP server and the web interface share one API, so agents and people get identical answers. It all runs on Edge infrastructure, with the training and indexing jobs on Edge Compute.
From four tabs to one question
Twenty thousand films indexed and scored in beta, built from an empty database at the start of an eight-month build, with the full launch in October 2026. Picking a film is a sentence rather than a search, and the list of what you meant to watch finally lives somewhere that remembers it. For Fi1m, three years of thinking became a product with two front doors: one for people, one for their agents, and a corpus underneath it that is now an asset in its own right.
Where to watch, and who to watch with
Handed to Edge Expert Services to run. On the roadmap after launch: availability across streaming services folded into the answer, shared lists and group decisions for households that can never agree, and television. Labs stays on as R&D partner.
Built for people and their agents.
Fi1m is the first Labs product designed from day one to be consumed over MCP as much as through a browser. It's also the largest corpus Labs has built from nothing: twenty thousand titles, resolved and indexed, with models trained on the result. That's the pattern we expect most consumer products to follow, and it's why the frontier report keeps coming back to it.
- Client
- Fi1m, film.one
- Sector
- Media & entertainment · Discovery platform and MCP
- Status
- Beta, launching October 2026
- Build
- Three years of groundwork, eight-month build
- Labs' role
- Report → concept and spec → build → handover
- Now run by
- Edge Expert Services
- Stack
- Corpus built from scratch: 20+ source ingestion and entity resolution · Fine-tuned review readers and score calibration · Semantic index over the catalogue · Explainable blended scoring · Conversational discovery · MCP server · Edge Compute, Storage and CDN
Give your product a second front door.
Your customers' agents are about to start using your service on their behalf. If you'd like that to work well, it's a Labs conversation.
Talk to Edge Labs