Question 1

Will it tick off a point I only mentioned in passing?

Accepted Answer

Yes, if you set the threshold low enough. The default 1.5 sits between "mentioned" and "discussed" and is adjustable from 0.5 to 2.5 on the page; at 1.0 a passing mention was enough to check a point off — saying "performers" before saying anything about tipping them. Each point returns a 0-3 score that is the probability-weighted level, so it can land between levels, and that threshold is the only knob you have.

Question 2

What does a full talk cost, and what makes the number move?

Accepted Answer

The README claims at most about $0.50 an hour at two calls a second, and less in practice because calls stop while you are silent and checked points are no longer sent. Its own replays: the 383-word example opening took 26 calls averaging 236 ms for $0.0011; a 1,581-word recorded talk took 106 calls averaging 252 ms for $0.0074. Both key paths pin Jev 1.13, follow the SDK's 10-second timeout with up to two retries honouring Retry-After, and stop on a 401.

Question 3

Can I trust it not to check off something I never actually said?

Accepted Answer

Reasonably, but the evidence is one dataset. Replaying a 383-word opening that covers the points out of order and never mentions the bar, that point never checked off and peaked at 0.03; on a longer 1,581-word talk the never-mentioned point peaked at 0.23. Because every open point is rescored on every call, order does not matter — but the score for a borderline point can drift as surrounding speech changes, and confidence will not warn you.

Question 4

Where do my API key and my transcript actually go?

Accepted Answer

An OpenRouter key goes from your browser straight to openrouter.ai. A TypeSafe key goes to server.js on your own machine, which passes it to api.typesafe.ai and neither stores nor logs it; put TYPESAFE_API_KEY in .env and the page never sees a key at all. A pasted key lives in session storage, or local storage if you tick "Remember on this device", until "Forget key". Every request also carries your transcript to that same service, and in Chrome speech recognition sends microphone audio to Google.

Question 5

Where does this design come from?

Accepted Answer

From Allie at TypeSafe, who built the original IntelliPrompter so she would not forget anything in the long list of announcements opening her stage shows, and demoed it at a TypeSafe Discord town hall. Running it on LLMs cost her roughly $40 an hour; the same design on Jev is at most about $0.50 an hour at two calls a second, and less in practice because calls stop while you are silent and checked points are no longer sent. This is an independent rebuild of her design as she described it.

Question 6

How does transcription work, and can I use it without a microphone?

Accepted Answer

Allie used ElevenLabs for live transcription; this page uses the browser's built-in speech recognition, which works in Chrome, Edge or Safari, and the shared logic lives in prompter.js. Start without microphone lets you type or paste what you said instead, which is also how npm run replay drives it from the terminal: node replay.js --every=10 --threshold=1.5, feeding a transcript 15 words at a time as live speech would arrive. Both key paths are used, TYPESAFE_API_KEY if set and OPENROUTER_API_KEY otherwise.

Question 7

How do I run it locally, and why would I?

Accepted Answer

Node 18 or newer and nothing to install: git clone, cd, optionally cp .env.example .env to add TYPESAFE_API_KEY and/or OPENROUTER_API_KEY, then npm start for http://127.0.0.1:8787/. The microphone needs localhost or HTTPS, so open the page through the server rather than as a file. The reason to run it is the TypeSafe key path: api.typesafe.ai refuses calls from web pages on CORS, so the bundled server.js relays them and never stores or logs the key.

Intelliprompter

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev guides

From the Jev community