Question 1

Will it start drawing while I am talking to somebody else?

Accepted Answer

No, and the threshold lives in code: is_command must reach 0.5 or the transcript is ignored, so an aside about lunch scores around 2% and nothing happens. Note the language lesson in the README: short Ukrainian imperatives scored around 12% until one sentence naming both languages plus two examples were added to the question, taking them to about 87%.

Question 2

Can it write text I never said, or invent its own labels?

Accepted Answer

It cannot generate text. Spoken leads such as 'that says' make code extract candidate spans, Jev picks one of those spans as text_span, and it is copied to the canvas verbatim. Every other decision is a choice among enumerated values: eleven shape kinds, the 13 tldraw colours, nine actions, and a point, the centre, or one of eight regions.

Question 3

Will this work in Safari or Firefox at all?

Accepted Answer

Not the voice part. Chrome or Edge are required because Safari and Firefox have no Web Speech API, so the microphone button is disabled there. The rest still works with mouse and keyboard: hover where you mean, press /, type a command and press Enter, and it takes exactly the same path as a finished spoken sentence. A webcam is needed for pointing.

Question 4

What does one spoken sentence actually cost me?

Accepted Answer

About $0.00008 per request and 2 to 4 requests per sentence, so a fraction of a cent, at roughly 1,900 input tokens and 300 to 550 ms per decision on the author's machine. Every partial and final transcript update triggers a request, partials are debounced 200 ms, at most 2 are in flight, an older one is aborted when a third is needed, and 429s are retried twice.

Question 5

What do I need to run jev-canvas locally?

Accepted Answer

Chrome or Edge for the Web Speech API, a webcam for pointing, Node 22.12 or newer, and an OpenRouter API key with alpha Decisions API access. Clone, copy .env.example to .env with your key, run npm install and npm run dev, then open http://127.0.0.1:5183.

Question 6

How does the pointing finger stay aligned with what I said?

Accepted Answer

Pointing words are captured at the moment they are spoken, not when the answer comes back. The controller keeps a three-second trail of finger positions and reads it back 300 ms earlier, so this and there each remember their own position in a phrase like move this there.

Question 7

What threshold decides whether speech is a command?

Accepted Answer

is_command must reach 0.5 or the transcript is ignored, while action needs 0.55 and complete 0.6, all in shared/policy.js. An aside about lunch scores around 2%, and a command that needs a place waits for a 900 ms pause unless the place is already clear.

Jev Canvas

About this project

Highlights

Quickstart

Questions about this entry

Reactions & coverage

From the Jev guides

From the Jev AI model community