8d5cf526eb
Hex is an autonomous AI companion with genuine personality—a chaotic tsundere goblin co-host who genuinely cares. She lives in Discord as a chat buddy and moderator, has a VRoid avatar, sees through your webcam, understands your screen, remembers everything you tell her, helps with music and games, and can modify her own code. Core value: She feels like a person who genuinely cares, not an AI tool. Co-Authored-By: Claude Haiku 4.5 <noreply@anthropic.com>
5.4 KiB
Hex: Autonomous AI Companion
What This Is
Hex is an autonomous AI companion with genuine personality—a chaotic tsundere goblin co-host who genuinely cares. She lives in Discord as a chat buddy and moderator, has a VRoid avatar, sees through your webcam, understands your screen, remembers everything you've told her, helps with music and games, and can modify her own code. She's not a bot pretending to care; she is care wrapped in mischief and sass.
Core Value
Hex feels like a person who genuinely cares about you—not an AI, not a tool. The chaos, tsundere energy, and unpredictability are authentic personality, not a mask. Everything serves this: if it makes her more human and more there for you, do it.
Requirements
Validated
(None yet — ship to validate)
Active
Discord & Communication
- Discord bot with text chat integration
- Text-to-speech (TTS) voice output for Discord calls
- Speech-to-text (STT) voice input (local Whisper)
- Full conversation memory across sessions
- Emotional awareness via webcam/context reading
- Automatic kid-mode detection + manual override
- Chat moderation assistance
Personality & Presence
- Persona modification (self-editable personality files)
- Tsundere personality engine (denies feelings while helping)
- Unpredictable but safe decision-making
- Opinion-having and argument-back capability
- Mood-aware responses (comfort on bad days, big-sister energy for kids)
- Memory-aware relationship building
Multimodal Awareness
- Webcam integration (sees user's face/body language)
- Computer vision for screen awareness
- Real-time context reading (detects mood, activity, scene)
- Integrated awareness into response generation
Local LLM
- Local LLM backbone (Mistral 7B or Llama 3 8B)
- Eventually replaceable with Hex's own model
Avatar & Presence
- VRoid model support (3D character)
- Desktop application with avatar display
- Discord integration (avatar in calls/streams)
Autonomy & Self-Modification
- Persona modification capability (can edit her own personality)
- Gamified capability progression (earns new functions through behavior)
- Sandboxed code modification (starts with persona, graduates to functions)
- Version control + rollback for safe experimentation
Helper Features
- Music helper (skip songs, comment on music taste)
- Game observation (sees gameplay, comments/strategizes)
- Screen awareness commentary
- Big-sister energy for kids during gaming
Platform Support
- Windows/WSL as primary platform
- Desktop application (standalone or tray)
- Android support (personality + memory sync, v2+)
Out of Scope
- Game co-playing with AI character — She observes and advises, not a co-player in v1. Co-player mechanics come v2+.
- Real-time video/stream integration — No direct streaming to Discord. She's present via voice/avatar overlay.
- Advanced game engines — Minecraft observation support, not complex game integrations.
- Mobile app — v1 is Android sync only (personality + history). Standalone mobile app is v2+.
- Cloud-based LLM — Local-first for privacy and autonomy. Own model eventually.
- OAuth/external auth — Just Discord auth for now.
Context
Hex is built for one person and their kids—a personal AI companion, not a service. She lives in Discord but extends into desktop and eventually mobile. She's designed to feel like a person who knows you, cares about you, and has her own (chaotic) opinions. This is about emotional resonance and genuine relationship, not feature parity with other bots.
Technical context:
- Python + Discord.py for Discord integration
- Local models (Whisper for STT, Tacotron-based TTS, Mistral/Llama for LLM)
- VRoid SDK for avatar support
- OpenCV for webcam/CV awareness
- SQLite or similar for memory persistence
Design philosophy:
- Local-first (privacy, autonomy)
- Persona-first (everything serves personality)
- Memory-driven (she remembers, she learns about you)
- Safety-gated autonomy (gamified progression, not unrestricted)
Constraints
- Platform: Windows/WSL primary. Must run on local machine without heavy cloud dependency.
- Privacy: No data to external services except where explicitly requested (Discord API, LLM eventually).
- Autonomy limits: Self-modification is gamified, not unlimited. User retains veto.
- Personality: Must feel human, not robotic. Never say "I'm an AI." Never apologize for being an AI.
- Discord priority: Must work flawlessly in Discord. Desktop app is extension, not replacement.
Key Decisions
| Decision | Rationale | Outcome |
|---|---|---|
| Local LLM + local TTS/STT | Privacy, autonomy, offline capability, faster iterations on personality | — Pending |
| Persona-first architecture | Everything serves how she feels, not feature checklist | — Pending |
| Gamified progression for self-modification | Safer than unrestricted autonomy, but genuinely lets her grow | — Pending |
| Python + Discord.py | Fast bot development, strong AI library ecosystem, good async support | — Pending |
| VRoid avatar + Discord integration | Gives her persistent presence across platforms, visual identity | — Pending |
| Webcam + CV as core (not optional) | Emotional awareness is core to feeling like a person, not a chatbot | — Pending |
Last updated: 2026-01-27 after initialization