In development · Amazon Developer Hackathon 2026
Trace watches the broadcast with you. It transcribes what is said, pulls out the people, organisations, places and things being named, and draws the connections between them on screen — live, as the story unfolds. News channels, conference talks, agency feeds — anything built around people talking.
STORY 1 OF 2
Listening…
The Fire TV screen, illustrated. Nothing is recorded in your living room — why that matters.
One graph per story
Run a whole evening through one graph and you get a hairball: the weather presenter ends up beside a foreign minister, because the only thing they share is having been said out loud in the same two hours. A day of conference sessions does the same thing.
So Trace watches for the moment the programme moves on. When a stretch of the broadcast starts naming people the current graph has never heard of, and keeps doing it, that is a new story — and it gets a graph of its own and a headline of its own.
The one being told now grows on screen. The ones already told stay on the remote, a press away.
Emissions accord signed in Geneva
0:00 · 12 ENTITIES
Rail strike ballot closes tonight
4:20 · 11 ENTITIES
Central bank holds rates
9:05 · 9 ENTITIES
Crew named for the next station flight
13:40 · 14 ENTITIES · ON SCREEN
How it works
Pick a channel on the TV, or have your phone send it any stream address. Then press Start. Nothing runs until you ask for it, and stopping actually tears the pipeline down.
Trace ingests the broadcast stream server-side and transcribes it with speaker separation. Your Fire TV never touches a microphone.
People, organisations, places, events and named objects are pulled from the transcript, along with the relationships stated between them. Repeat mentions merge into one node.
Nodes spring in as they are named, edges draw between them, and the layout settles. Size is how central an entity is; colour is what kind of thing it is — person, organisation, place, event.
Move the remote to a node and press OK. You get a summary grounded in what was actually said — and, on request, a web-sourced profile with citations.
Save the graph at the end of the segment. Come back to it later, or open it on the web to rename, merge and prune what the pipeline got wrong.
What you get
Every on-screen summary comes from the transcript you just heard. When you ask for more, Trace searches the web and shows you the sources — it never answers from a model’s memory. Live talk about real people is exactly where a model invents them.
Pause the graph to read it without the layout moving under you. Stop the trace to end the session outright. The two are different buttons because one of them costs money and the other does not.
Rename an entity, merge two nodes that are the same person, delete noise, pin what matters. Your edits are stored as a layer over the original capture, so the raw graph is never lost.
No Fire TV has a touchscreen, so nothing here pretends otherwise. Focus moves between nodes on the D-pad, type is readable from the sofa, and every target is reachable in a few presses.
Export to JSON, PNG or GraphML. Open a saved map in the web app for the full picture, with search across the transcript that produced it.
The graph renderer is being extracted into a standalone MIT-licensed package, so anyone building for Vega OS can draw a live force graph without starting over.
What you can point it at
Trace works on people being named out loud. Nothing in the pipeline is specific to news — news simply happens to be where you most often hear a name you do not know. Pick from the channel list on the TV, or send any stream address from your phone.
Netflix, Prime Video, YouTube
They hand out no stream address, so there is nothing to send. This one is not a feature we have not built yet — it cannot be built.
Drama and film
The graph would be of invented people, and looking them up on the web returns somebody real. Wrong on both ends.
Sport
Names arrive faster than anywhere else and the connections between them are never said out loud.
Music, or anything barely spoken
Nothing is named, so nothing is drawn.
Privacy
This is an architectural fact, not a policy promise. The Fire TV app never captures audio — not from the microphone, not from the system mixer. There is no capture path in the app at all.
What gets transcribed is the broadcast stream itself, ingested server-side from the same source your TV is playing. The device receives a graph back. It never sends one up.
One useful consequence: because the channel is ingested once and served to everyone watching it, your viewing does not create a recording that belongs to you. There is nothing personal in the pipeline to leak.
Pricing
A channel is transcribed once and served to everyone watching it, which is what makes a genuinely free tier possible. The paid plans cover the things that cost per person: web lookups, storage, and streams only you are watching.
$0forever
Watch a catalog channel and see the graph build itself.
$4.99per month
Keep everything you build, and find out who these people are.
$12.99per month
Point Trace at your own stream, and pull the data back out.
Prices are indicative while Trace is in development, and nothing is charged yet. Plans will be purchasable through the Amazon Appstore on the TV and by card on the web.
It is being built for the Amazon Developer Hackathon, deadline 23 October 2026. Leave an address and you will hear when it lands on the Fire TV Appstore.
One email when it ships. No drip campaign, and you can leave at any time.