YURIOS // LAB
The Codex → Appendices Appendix B

Tools and Stacks

Curated short lists, by layer

Curated short lists, organised by layer. Confirm current versions before committing — this list ages fast.

Inference (hosted)

  • Anthropic Claude (Opus/Sonnet/Haiku 4.x) — current best persona stability per dollar at the top.
  • OpenAI GPT-5 family — strong default; native realtime + tool stack.
  • Google Gemini 2.x — long context, good multimodal.
  • Together / Fireworks / Replicate / DeepInfra / Modal — hosted open-weight inference.
  • Groq / Cerebras — sub-100ms TTFT for small models.

Inference (local / self-host)

  • vLLM — server-side default for open models.
  • SGLang — competitive vLLM alternative.
  • llama.cpp / llamafile / Ollama — laptop-grade local.
  • MLX — Apple Silicon local.
  • TensorRT-LLM — fastest if you've committed to Nvidia.

Models (open) 2026 short list

  • Llama 4 family (Meta)
  • Qwen 3 family (Alibaba)
  • DeepSeek V3 / R1
  • Mistral Large 3
  • Gemma 3 (Google)
  • Small specialists: Mistral Small, Qwen 2.5-7B, Gemma 3-9B, Phi-3.5

Memory / RAG

  • pgvector — Postgres + vectors; the default starter.
  • Qdrant / Weaviate / Chroma — dedicated vector DBs.
  • Mem0 — opinionated memory framework.
  • Letta (née MemGPT) — agent-with-self-managed-context.
  • Zep / Graphiti — temporal + graph memory.
  • GraphRAG (Microsoft) — graph over a doc corpus.

Agent / orchestration

  • MCP servers (Anthropic protocol) — tool surface standard.
  • LangGraph — explicit graph orchestration; production-grade default.
  • CrewAI / AutoGen — multi-agent (use sparingly).
  • Claude Agent SDK / OpenAI Agents SDK / Strands (AWS) — frontier-lab / hyperscaler offerings.
  • Mastra — TypeScript-native, Next.js-friendly.
  • Pydantic AI — typed, production Python.
  • ElizaOS — TypeScript agent framework with character JSON, multi-agent, and out-of-box Discord/Telegram/X/Farcaster/Bluesky/Slack connectors. Web3-flavored but useful for the social-presence layer (→ ch. 02 §5).
  • Pipecat / LiveKit Agents — real-time voice agents.

Voice — TTS

  • ElevenLabs — closed, prosody king.
  • Hume EVI — voice + emotion.
  • OpenAI Realtime API — integrated stack.
  • XTTS-v2, StyleTTS2, Kokoro, Orpheus — open.
  • Sesame, Cartesia — newer, watch list.

Voice — STT

  • Whisper / faster-whisper / Distil-Whisper — open default.
  • Nvidia Parakeet / Canary — faster.
  • Deepgram / AssemblyAI — hosted.

Avatars

  • Live2D Cubism Editor + SDK — 2D rigged.
  • VRoid Studio — easy VRM creation.
  • Blender + VRM plugin — serious VRM.
  • three-vrm — browser VRM runtime.
  • UniVRM — Unity VRM.
  • VRM4U — Unreal Engine VRM with built-in VMC blueprints.

Avatar control protocols & host runtimes

The AI-brain → avatar pipeline. See ch. 25 for the architecture and trade-offs.

  • VMC Protocol — OSC-over-UDP standard for VRM control (MIT). Ports 39539/39540. Senders: VSeeFace, Waidayo, MocapForAll, TDPT. Receivers: Warudo, VNyan, EVMC4U (Unity), VRM4U (UE5), VMC4B (Blender).
  • VTube Studio API — WebSocket/JSON plugin API for Live2D. VTubeStudioJS (TS/JS), Python clients, @hkopenai/vtube-studio-plugin-mcp-server (MCP).
  • ARKit blendshape vocabulary (52) — the de facto facial expression naming standard. Emit ARKit names; most hosts understand them.
  • Warudo — proprietary freeware 3D VTuber app with Steam Workshop plugin ecosystem + C# SDK. VMC receiver.
  • VNyan — freeware; node-based visual scripting; VMC receiver; node-graph friendly to AI integrations.
  • VSeeFace — freeware; VMC sender + receiver; classic face-tracking → VRM.
  • VTube Studio — Live2D incumbent; one-time paid; WebSocket plugin API.
  • AITuberKit — open-source Next.js + TS toolkit for AI VTubers; VRM + Live2D + Motion PNG; YouTube Live ingestion built in. Closest stack-match to this project's existing repos.

Image generation

  • ComfyUI — open default, power-user.
  • A1111 — older but ubiquitous.
  • SDXL / Flux / SD3 — base models.
  • NovelAI — closed reference for anime consistency.

Safety / moderation

  • Llama Guard 3 / 4 — open classifier.
  • NeMo Guardrails — programmable.
  • Granite Guardian (IBM) — jailbreak focus.
  • OpenAI Moderation / Anthropic safety classifiers — bundled.

Frontends / shells

  • SillyTavern — the RP power-user UI.
  • Open WebUI — clean LLM web UI.
  • AIRI — 40k-star self-hosted multi-platform AI waifu (VRM/Live2D, voice, memory, plays games). The closest existing implementation of the agentic-waifu spec (→ ch. 02 §5).
  • Soul of Waifu — full-stack desktop reference.
  • Open-LLM-VTuber — voice + Live2D cross-platform.
  • Electron / Tauri — desktop wrappers.
  • Next.js / SvelteKit / SolidStart — web frameworks.
  • Three.js / Babylon.js — browser 3D.

Distribution

  • GitHub — code + releases.
  • HuggingFace — models, adapters, datasets, Spaces.
  • Chub.ai / CharacterHub — character cards.
  • AICharacterCards.com — character cards.
  • JanitorAI — RP frontend + audience.
  • VRChat — VRM avatar distribution = marketing.

Creator economy

  • Stripe — payments.
  • Ko-fi — donations + small shop.
  • Patreon — recurring patronage.
  • Buttondown / Ghost / Substack — newsletter.
  • Gumroad / Lemonsqueezy — digital product sales.
  • Discord — community.

Compliance / safety vendors

  • Yoti / Persona / AU10TIX — age verification.
  • C2PA / Truepic — content provenance.