Meeting copilot · runs on your own computer

Lost the thread? Press one key.

ContextLens listens to your meeting. Press, and a card shows what was said, who said it, and what changed, quoted straight from the conversation.

Desktop and smart glasses · Open source · Nothing leaves your machine

Four people in a call, shown as a bear, a robot, a fox and a cat. Next to them, a ContextLens card that appeared on its own says the beta now ships Monday, quotes Maya, and marks that Monday replaced the Friday date she gave first.

How it works

  1. 1

    It listens

    Transcribes the meeting and labels who is speaking, all on your machine.

  2. 2

    You press

    One hotkey on desktop, a tap on glasses. When something changes, it speaks up on its own.

  3. 3

    You read one card

    Short, with quotes and who said them. If nobody said it, the card says so.

Two ways to run it

Built for people who find live conversation hard to track, including introverted and neurodivergent people.

The ContextLens card floating over a video call: the beta moved from Friday to Monday, with the quotes and who said them.

Desktop

Tray app, global hotkey, always-on-top card.

See the desktop app

The ContextLens card floating in a dim meeting room as the glasses show it: answered, the beta moved from Friday to Monday, with Maya’s quotes and times, and Press, Pause, Resume, Webcam and Delete buttons.

Smart glasses

The Unity client. Tap to press.

See the glasses client

Try a card

Pick a question. The card answers from the conversation and marks the lines it used.

Who said what
The quote and the speaker.
What changed
Friday became Monday, and who moved it.
Not sure
It says “not enough information” instead of guessing.
Planning sync, 4 minutes in
  1. t1MayaLet’s ship the beta on Friday.
  2. t2JordanDesign still owes the onboarding screens.
  3. t3JordanThen QA needs two days, minimum.
  4. t4MayaOkay, make that Monday.
FridayMonday
Who
Maya
Where
t1, t4
Read as
An update, not a contradiction

Tested on long real meetings

161 held-out cases from QMSum and AMI, with conflicts 20 to 60 turns apart.

Conflicts caught
87%Single agent 0%
Found evidence 10+ turns back
60%Single agent 0%
Answers graded correct
65%Single agent 35%
Contradiction F1 on DECODE
0.913
Baseline 0.753, on 1,000 pairs
False alerts
Halved
Against that baseline
Time per check
45 ms
On every press

See the full evaluation

Under the hood

A small local model, and the system around it that keeps it honest.

  • Streaming speech

    whisper.cpp, C++, gRPC.

  • Speaker labels

    Who is talking.

  • Memory

    Builds as the meeting goes.

  • Contradiction check

    ModernBERT NLI and a fine-tuned DeBERTa, pair by pair, in code.

  • Evidence first

    Code picks and cites the turns; the model only words the answer.

  • Hybrid retrieval

    BM25, vectors, a bge cross-encoder reranker.

  • Evidence guard

    Drops claims no earlier turn backs.

  • Local model

    4B, through Ollama, on your machine.

See how it’s built

See it in 80 seconds

No install. A short meeting plays in your browser and the cards appear on their own.