talkweave. augment the conversation. never interfere.

a black screen. a growing transcript. one key that means don't let me lose that.

talkweave sits on the table between two people talking. it never talks back. this is what it looks like, playing out below, right now.

0:00 · live

left okay so what is this thing actually
right a transcript that grows in the middle of a black screen. your words go left, mine go right. that's all you see.
left and while I'm talking, if you make a point I want to come back to…
right one key lights the point amber and drops it in my sidebar, so it waits there until it's my turn to talk. ↪ click this line to flag it
left so nothing gets lost between my sentence and your reply.
right that's the whole point. augment the conversation, never interfere with it. ↪ click this line to flag it
the gesture

flagging takes one key. it never stops the conversation to do it.

while the other person talks, you flag. while you talk, they flag. by the time either of you gets a turn, the point is already sitting there, exact, amber, waiting.

←→

who is speaking right now

left shiftright shift

left shift flags a point for the left person, right shift for the right, pressed alone

space

the listener flags whatever was just said

double-click a word

flag that exact claim, nothing more, nothing less

F

fact-check the flagged point, locally

try the last one

remote work cut commercial real estate demand by half.

double-click any word above. in the app, this is exactly how a claim gets flagged mid-sentence, without breaking your flow.
the fact-check

quick answer in seconds. a deeper model keeps digging behind it.

press F on a flagged point and a small local model answers first. a slower, more careful model keeps working after that, with an eta, and it says so plainly if the two disagree.

flagged claim: “the policy passed in 2020.”

the record

every flagged point sits on a timeline you can walk back through.

one-word notes next to a flag. a summary every fifteen minutes. and underneath all of it, an audio timeline you can click to jump back to the moment something was actually said. illustrative below, not a real recording.

jumped to 7:15, right where the flag went amber

where it runs

on your machine. only your machine.

no cloud.
no account.
no telemetry.
works without any llm at all.
ollama is optional, only for claim extraction and fact-check.▏

running it

three commands. no build step you have to think about.

$ npm install && npm run build && node server/server.mjs
→ http://127.0.0.1:8990

ollama is optional. install it and pull a small and a larger model to turn on claim extraction and fact-check. everything else, the transcript, flagging, notes, summaries, the audio timeline, runs without it.

honestly, right now

live transcription shipped: it runs against a local qwen3-asr model, not whisper, with the first word on screen in about 1.4 seconds, measured, while you're still speaking. the flagging, the sidebars, the fact-check and the timeline are real and working today, alongside it.