Open Annie

An open-source 3D character you talk to in the browser from rehan-remade. One model talks: GPT-Live-1, a full-duplex speech model, speaks over WebRTC. A second model moves her: Jev reads the live transcript and picks her face, gestures and body language in about 120 ms.

While you talk, Jev answers one question: which face a listening friend would make. While she talks, transcript fragments trigger at most four asks a turn, each asking five questions in one request: committing to an action, which clip, which reaction, which face, and how much energy.

Yes. From a clone, python3 -m http.server 8080 serves a scripted preview with Annie's lines from Kokoro-82M and decisions from the local rule floor, plus a replay of the recorded live session with its real Jev decisions. Only live conversation needs keys.

An OpenAI key with GPT-Live access and a TypeSafe key, copied into .env; then build the broker virtual environment on Python 3.12, run broker local.py on port 8787, open the stage with live=1, and allow the microphone.

They come from the project's own recordings: 51 decisions at p50 118 ms and p90 151 ms, and $0.0026 of Jev for one 97-second conversation. The README also measures Jev upstream at p50 108 ms over 20 warm sequential requests and notes the broker adds about 1 ms. These are single-project measurements, not third-party benchmarks.

Code keeps control: Jev chooses only from installed clips, every answer passes a confidence gate before it can act, actions like dancing require a spoken commitment, and a refusal vetoes motion. Any Jev miss falls back to a local rule floor so she never waits.

rehan-remade, with code under Apache-2.0; assets carry their own licences recorded per file and checked in CI, including pixiv sample-model terms for two avatars and a CC0 motion pack. The digest shows 41 stars, 11 forks, and creation and push on 2026-10-07.