Send it a meeting link. It joins as its own participant, answers out loud, and draws on a whiteboard.
…so are we shipping the migration this week?
To be in a video call you need to be two devices: a camera and a microphone. Everything else is detail. The AI half is commodity. The device half is where the work is.
Chrome plays the meeting into a sink that isn't a speaker. One parec records its monitor, and every participant is mixed for free.
The monitor stream becomes text on the host, where it keeps its GPU acceleration. Base English model, no timestamps.
A small non-reasoning model answers in a third of a second and tags its own expression inline, e.g. [thinking].
The reply becomes PCM through mlx-audio, a voice that doesn't sound like a satnav, and its amplitude drives the mouth.
The avatar is drawn into a canvas at about a millisecond a frame, then handed to Meet as a camera track.
Patched before Meet's own JavaScript runs. Meet enumerates a camera and a microphone and cannot tell they aren't hardware.
Chrome routes decoded remote audio to the output device and never hands it back to WebAudio, so intercepting tracks in the page is a dead end: it attaches happily and delivers silence. Recording the sink monitor instead also means Tekki can't hear itself: its voice leaves by the microphone track and never reaches a speaker.
Measured over the same three turns: qwen3:4b took 23s to answer cold, then returned empty replies at 38s and 40s. On a reasoning model the token budget covers the scratchpad first, so a cap sized for two spoken sentences gets spent thinking.
| turn | qwen3:4b | 3B, no reasoning |
|---|---|---|
| first, cold | 23.4s | 10.1s |
| second | 38.1s · empty | 0.32s |
| third | 40.8s · empty | 0.32s |
Not to save effort, and not as a placeholder for something photoreal later.
Anyone glancing at the call can see there's a bot in it. That's a promise the consent section makes, and an antenna is what keeps it.
Any face implies an age, a gender and an ethnicity, and every viewer reads something into those. A robot sidesteps the question entirely.
A drawing that blinks and reacts reads as charming. A nearly-human face that blinks slightly wrong reads as a corpse.
And it stays cheap: arcs and rounded rectangles, about a millisecond a frame, where a diffusion model needs hundreds. That's the whole reason local realtime is possible at all.
Every tile is the live renderer, not a recording, the same drawAvatar that feeds the camera track in a call.
~1ms / frameTutor mode
Send a picture of what you are stuck on, then a meeting link. It works out a running order and teaches it on a whiteboard it draws itself.
The one you're stuck on. A local vision model reads it into a topic, a formula, the key facts and a worked example.
It joins, turns its camera on, and shares a whiteboard of its own. Meet never opens a screen picker, because it is never asked.
It stops mid-sentence, answers the question, and picks up at the step it left off. The page is the syllabus, so it answers from the book rather than from memory.
The model never draws. It picks a shape by name and every coordinate is computed in code, so a thin answer becomes a thinner diagram rather than a broken one.
A supervisor takes the message and forks a join session per conversation, so several chats can each hold their own meeting, one call per profile. Telegram is for when you are away from the machine; the Mac app speaks to the same supervisor directly, and the token is optional.
In the call you can also just ask out loud. “Draw how that works” and “show me the difference” both land, and it will stop mid-sentence to answer a question.
Mac app
A small Mac app that checks what Tekki needs, starts it, and drives a call from a window: paste a link, say something, teach a page. It configures the same thing the commands below do; it does not replace them.
Tekki 1.0
macOS 14 or later · Apple silicon or Intel · 441 kB
Download for MacRelease notes and older versionsIt isn’t notarised. macOS will refuse to open it the first time, because it came from the internet and there is no paid developer identity signing it. Right-click the app and choose Open, then Open again. If you would rather not take that on trust, build it from source instead. Same app, no warning.
The app still needs the repo cloned, Docker running and the three local model services up. It checks for each and tells you which is missing. Everything below is what it is checking.
Needs Docker, and a Mac or Linux box with enough memory to hold a 3B model. No account to create, no key to buy, no endpoint to trust.
Ollama for the thinking, whisper for the listening, Kokoro for the voice.
ollama pull llama3.2:3bbrew install whisper-cppmkdir -p models && curl -L -o models/ggml-base.en.bin \ https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.en.bin # Kokoro, for a voice that doesn't sound like a satnavuv venv --python 3.12 .venvuv pip install --python .venv/bin/python mlx-audio "misaki[en]" fastapi uvicornThey stay on the host, not in the container, so they keep their GPU acceleration.
ollama servewhisper-server --model models/ggml-base.en.bin --port 8178 --host 127.0.0.1 --no-timestamps.venv/bin/python tools/kokoro_server.py --port 8179Use a throwaway. It appears in calls under that account's name. Nothing in the repo reads, stores or transmits a password.
cd bot && ./signin.sh # opens a browser at localhost:6080; sign in by handOnly if you want to drive it from your phone. The Mac app talks to the bot directly and needs no token. Otherwise: talk to @BotFather, send /newbot, and put the token in bot/.env.
TELEGRAM_TOKEN=123456789:AA...Message your bot /start, then send it a Meet link. Admit it when it knocks.
cd bot && docker compose upHonest section
Somebody has to admit it.
No diarisation. Everyone is one voice to it.
It can draw on a board of its own, but a share, a slide or a face is invisible to it. It reads a page you photograph, not one you hold up.
The board is only as good as the model behind it. On a subject a small local model doesn't know, it will draw a confident, wrong diagram.
Google refuses automated anonymous clients, so it needs its own account.
OAuth grants API scopes; there is no Meet API for joining a call as somebody else. That door doesn't exist.
It ignores the noises whisper invents out of silence, and acknowledgements while it's teaching. Everything else it hears, it answers: fine one-to-one, rude in a group.
Consent
Tekki joins under its own name, renders as an obvious cartoon robot, and doesn't attempt to look human. Anyone can see there's a bot in the call. Any feature that made it harder to notice would be the wrong feature.
Also in the repo
SwiftUI, putting the avatar in a window for OBS to capture and speaking through BlackHole. It works, and it's how the avatar, brain and voice were built. Becoming a device inside the page removed two prerequisites and all the manual setup.
swift run understudy-window
swift run understudy-brain
swift run understudy-renderClone it, pull a 3B model, and message a bot a link. Everything it hears stays on the machine it runs on.