AI VTuber - 24/7 Autonomous Live Streaming System - Sell Project

A complete, working codebase for an AI-driven VTuber that streams continuously on YouTube and replies to live chat by voice, in real time, through a 3D avatar. YouTube chat -> moderation -> LLM -> text-to-speech -> 3D avatar -> OBS -> live stream This is sold as a CODEBASE, not as a running channel or a revenue stream. Please read the honest disclosure below before contacting me. === MEASURED SCALE (counted, not estimated) === Python: 3,006 lines across 13 files. Avatar renderer (browser source): 453 lines using Three.js and three-vrm. Configuration: 231 lines of YAML across 12 sections. Operations manual: 408 lines. Runtime dependencies: only 7 (aiohttp, anthropic, chat-downloader, numpy, pillow, pytchat, PyYAML). All 13 Python files pass a syntax check. === WHAT IT DOES === Reads YouTube live chat (pytchat by default, chat-downloader as fallback). Moderates BOTH the incoming comment and the generated reply against configurable word lists. Generates replies with Google Gemini (free tier), Anthropic Claude, or a local Ollama model - switchable in config. Hybrid routing sends small talk to a cheap model and keeps the stronger model for when it matters. A hard cost ceiling in config stops a 24-hour stream from running away with your API budget. Speaks through VOICEVOX. Drives a VRM avatar in the browser with lip-sync, served to OBS as a browser source. An "idle bank" of topics fills silence so there is no dead air when chat is quiet. A built-in self-diagnostic checks LLM, voice, overlay and moderation one by one and reports in plain language what is missing before you go live. Includes a VRM recolouring utility to restyle an avatar's textures without a 3D editor. === WHY THIS HAS VALUE === The hard part of an AI VTuber is not the LLM call. It is keeping chat ingestion alive for hours without being rate-limited or silently dropping messages; moderating in BOTH directions so the character cannot be baited into ending your channel; keeping audio, lip-sync and the avatar in step when replies arrive at irregular intervals; holding API cost to something survivable when the stream never stops; and filling silence without producing the repetitive output that platforms refuse to monetise. This codebase is already past all of those points. The operations manual documents the platform rules that govern them - AI-content disclosure, TTS credit requirements, and the "mass-produced / repetitive content" monetisation policy - with the specific design choice made for each. === HONEST DISCLOSURE === I never went live with it. Self-diagnostics passed and the system was verified on my own machine in August 2026, but the final step - pasting a YouTube stream key into the config - was never taken. There is no channel, no audience, no watch time and no revenue. I am not selling a business. I am selling the implementation. The last full runtime verification was August 2026; dependencies are pinned (a frozen requirements file from September 2026 is included), but I have not re-run the system since. Expect to spend time on environment setup. === KNOWN LIMITATIONS (all disclosed) === 1. Text-to-speech is Japanese-only. VOICEVOX is a Japanese TTS engine. To stream in another language you must replace the voice layer. It is isolated in one module (src/voice.py), but this is real work, not a config change. 2. The avatar model is NOT included. The VRM file I developed against has no license documentation I can verify, so I will not transfer it. Buyers should create their own with VRoid Studio (free) or use a model licensed for commercial use. Everything that loads and animates a VRM is included and works with any standard VRM. 3. Designed for WSL2 (Ubuntu) with OBS on Windows. It will run on Linux generally, but that is the configuration that was tested. 4. Never run against a live audience. Behaviour under a fast-moving chat is unproven. 5. Running unattended is a risk you accept. The manual recommends reviewing archives daily for the first week. The moderation layer is good, not perfect. === REQUIREMENTS TO RUN IT === Python 3.11+, OBS Studio, a VOICEVOX installation, a VRM avatar model, and an API key for Gemini (free tier is enough to start), Claude, or a local Ollama install. === TERMS === Payment via escrow, Wise, or Stripe. No earn-outs or revenue-share arrangements. Communication in English is BY TEXT ONLY. I do not take voice or video calls. I answer written questions in detail, usually within 24 hours. Handover: repository transfer plus a written walkthrough. I will answer questions by text for 14 days after transfer at no charge.