15 KiB
15 KiB
MoAI - Master of AIs
A multi-AI collaborative brainstorming platform where multiple AI models (Claude, GPT, Gemini, etc.) discuss topics together, see each other's responses, and work toward consensus.
Domain: moai.dk (planned)
Concept
Instead of asking one AI a question, ask a "team" of AIs. They each provide initial thoughts, then engage in a structured discussion where each can see and respond to the others' contributions. The result is a richer, more diverse exploration of ideas that gets summarized into actionable markdown documents.
Core Features
Projects/Context
- Projects are containers for related discussions
- Each project maintains its own context and conversation history
- Projects have configurable model teams (e.g., Claude + GPT + Gemini)
- CRUD operations for project management
- Projects persist and can be revisited/continued
Two Question Modes
1. Open Question (/open)
- Question sent to ALL models in parallel
- Each model responds independently without seeing others
- Fast way to get diverse initial perspectives
- Results displayed together for comparison
2. Discussion Mode (/discuss [rounds])
- Sequential turn-based discussion
- Each model sees the full conversation history including other models' responses
- Models can agree, disagree, build upon, or challenge each other
- Continues until:
- Consensus is reached (detected by system)
- Round limit is hit (user-defined, e.g.,
/discuss 5) - User manually stops (
/stop)
Iteration Flow
┌─────────────────────────────────────────────────────────────┐
│ /open "Initial question" │
│ ↓ │
│ [Parallel responses from all models] │
│ ↓ │
│ /discuss 5 │
│ ↓ │
│ [5 rounds of sequential discussion] │
│ ↓ │
│ /consensus │
│ ↓ │
│ [AI-generated summary of agreements/disagreements] │
│ ↓ │
│ /open "Follow-up question" (includes consensus as context) │
│ ↓ │
│ [Cycle continues...] │
│ ↓ │
│ /export → Final markdown document │
└─────────────────────────────────────────────────────────────┘
Direct Model Interaction
@claude <message>- Direct a question/comment to a specific model- Other models see this exchange as part of the context
- Useful for drilling down on a specific model's point
Architecture
┌────────────────────────────────────────────────────────────────┐
│ Web UI (Phase 2) │
│ - Project CRUD, settings, model config │
│ - Full discussion timeline view │
│ - Export/download markdown │
│ - Rich text, syntax highlighting │
│ - Drag-drop, edit, reorder │
└────────────────────────────────────────────────────────────────┘
│
▼
┌────────────────────────────────────────────────────────────────┐
│ Python Backend (FastAPI) │
│ │
│ ┌──────────────────────────────────────────────────────────┐ │
│ │ API Layer │ │
│ │ /projects, /discussions, /messages, /export │ │
│ └──────────────────────────────────────────────────────────┘ │
│ │ │
│ ┌─────────────┐ ┌─────────────┐ ┌──────────────────────┐ │
│ │ Projects │ │ Discussions │ │ AI Orchestrator │ │
│ │ Service │ │ Service │ │ │ │
│ │ │ │ │ │ - Round management │ │
│ │ - CRUD │ │ - Open mode │ │ - Context building │ │
│ │ - Settings │ │ - Discuss │ │ - Consensus detect │ │
│ │ - Export │ │ - History │ │ - Model routing │ │
│ └─────────────┘ └─────────────┘ └──────────────────────┘ │
│ │ │
│ ┌───────────────┴───────────────┐ │
│ │ Database (SQLite) │ │
│ │ projects, discussions, │ │
│ │ messages, consensus │ │
│ └───────────────────────────────┘ │
└────────────────────────────────────────────────────────────────┘
│ │
▼ ▼
┌──────────────────┐ ┌─────────────────────────┐
│ Telegram Bot │ │ Requesty / OpenRouter │
│ (thin client) │ │ │
│ │ │ - Claude (Anthropic) │
│ - Commands only │ │ - GPT (OpenAI) │
│ - No state │ │ - Gemini (Google) │
│ - Calls API │ │ - Llama, Mistral, etc │
└──────────────────┘ └─────────────────────────┘
Data Model
Project
├── id: uuid
├── name: string
├── created_at: timestamp
├── updated_at: timestamp
├── models: ["claude-opus-4", "gpt-4o", "gemini-2-pro"]
├── settings:
│ ├── default_rounds: int (default: 3)
│ ├── consensus_threshold: float (optional, for auto-detect)
│ └── system_prompt_override: string (optional)
└── discussions: Discussion[]
Discussion
├── id: uuid
├── project_id: uuid (FK)
├── question: string
├── type: "open" | "discuss"
├── status: "active" | "completed"
├── created_at: timestamp
├── rounds: Round[]
└── consensus: Consensus (nullable)
Round
├── id: uuid
├── discussion_id: uuid (FK)
├── round_number: int
├── type: "parallel" | "sequential"
└── messages: Message[]
Message
├── id: uuid
├── round_id: uuid (FK)
├── model: string ("claude", "gpt", "gemini")
├── content: text
├── timestamp: timestamp
└── is_direct: boolean (true if @mentioned)
Consensus
├── id: uuid
├── discussion_id: uuid (FK)
├── agreements: string[] (bullet points)
├── disagreements: [{topic, positions: {model: position}}]
├── generated_at: timestamp
└── generated_by: string (which model summarized)
Telegram Commands (Phase 1)
Project Management
/projects - List all projects
/project new "Name" - Create new project
/project select <id|name> - Switch to project
/project delete <id> - Delete project
/project models claude,gpt - Set models for current project
/project info - Show current project details
Discussion
/open <question> - Ask all models (parallel)
/discuss [rounds] - Start sequential discussion (default: 3 rounds)
/next - Trigger next round manually
/stop - Stop current discussion
@claude <message> - Direct message to specific model
@gpt <message>
@gemini <message>
Output
/consensus - Generate consensus summary
/summarize - Summarize current discussion
/export - Export full project as markdown
/export discussion - Export current discussion only
Utility
/status - Show current project/discussion state
/history [n] - Show last n messages
/clear - Clear current discussion (confirm required)
/help - Show commands
System Prompts
Base System Prompt (all models)
You are participating in a roundtable discussion with other AI models.
Other participants: {list of models}
Current topic: {project name}
Question: {current question}
Guidelines:
- Be concise but substantive
- You can agree, disagree, or build upon others' points
- Reference other models by name when responding to their points
- Focus on practical, actionable insights
- If you reach agreement with others, state it clearly
Consensus Detection Prompt
Review the discussion above. Identify:
1. Points where all models agree
2. Points where models disagree (list each position)
3. Open questions that weren't resolved
Format as structured JSON.
Export Format
# {Project Name}
**Date:** {date}
**Models:** Claude Opus 4, GPT-4o, Gemini 2 Pro
**Discussions:** {count}
---
## Discussion 1: {Question}
### Initial Responses (Open)
**Claude:**
> {response}
**GPT:**
> {response}
**Gemini:**
> {response}
### Discussion (5 rounds)
**Round 1:**
**Claude:** {response}
**GPT:** {response}
**Gemini:** {response}
**Round 2:**
...
### Consensus
**Agreements:**
- {point 1}
- {point 2}
**Disagreements:**
- **Topic:** {topic}
- Claude: {position}
- GPT: {position}
**Open Questions:**
- {question}
---
## Discussion 2: {Follow-up Question}
...
---
## Final Summary
{AI-generated executive summary of all discussions}
Phase 1: Telegram POC
Scope
- Single-user (your Telegram only)
- SQLite database
- Basic project CRUD via commands
- Open and discuss modes working
- Export to markdown (sent as document)
- Requesty/OpenRouter integration
Tech Stack
- Python 3.11+
- python-telegram-bot (async)
- FastAPI (internal API, optional for phase 1)
- SQLite + SQLAlchemy
- httpx for API calls
- Requesty or OpenRouter for model routing
Coding Standards
- Formatting/Linting: ruff (line length 100)
- Testing: pytest (80%+ coverage on core logic)
- Type hints: Required on all public functions
- Docstrings: Required on modules and classes
- Logging: Use
loggingmodule, noprint()in production - Pre-commit: Enforce ruff and pytest before commits
- Dependencies: No pinned versions unless required (e.g.,
fastapinotfastapi==0.109.0) - prefer latest for security
File Structure
moai/
├── src/
│ └── moai/
│ ├── __init__.py
│ ├── bot/
│ │ ├── __init__.py
│ │ ├── main.py # Bot entry point
│ │ ├── handlers/
│ │ │ ├── __init__.py
│ │ │ ├── projects.py # /project commands
│ │ │ ├── discussion.py # /open, /discuss, @mentions
│ │ │ └── export.py # /export, /consensus
│ │ └── middleware.py # Auth, logging
│ └── core/
│ ├── __init__.py
│ ├── models.py # SQLAlchemy models
│ ├── database.py # DB connection
│ ├── orchestrator.py # AI round management
│ ├── ai_client.py # Requesty/OpenRouter wrapper
│ └── exporter.py # Markdown generation
├── tests/
│ ├── __init__.py
│ ├── test_models.py
│ ├── test_orchestrator.py
│ └── test_handlers.py
├── pyproject.toml # Project config, dependencies, ruff settings
├── .pre-commit-config.yaml
└── README.md
MVP Milestones
- M1: Bot responds to /help, /status
- M2: Project CRUD working
- M3: Single model Q&A working (no discussion yet)
- M4: Open mode (parallel) working with multiple models
- M5: Discuss mode (sequential) working
- M6: Consensus generation working
- M7: Export to markdown working
- M8: @mention direct messages working
Phase 2: Web UI
Additional Features
- Visual timeline of discussions
- Project dashboard
- Model performance/cost tracking
- Edit/delete individual messages
- Branching discussions (explore alternatives)
- Share/collaborate (multi-user)
- Templates for common brainstorm types
Tech Stack Options
- Frontend: Vue 3 + Vite or React + Next.js
- Backend: FastAPI (expand from Phase 1)
- Database: PostgreSQL (upgrade from SQLite)
- Auth: Simple token or OAuth
- Hosting: Dockge/Docker on homelab
Cost Considerations
Each discussion round involves multiple API calls:
- Open question: N calls (one per model)
- Discussion round: N calls
- Consensus: 1 call
Example: 3 models, 1 open + 5 discuss rounds + consensus = 3 + 15 + 1 = 19 API calls
Use Requesty/OpenRouter for:
- Single API key for all models
- Cost tracking
- Fallback routing
- Rate limit handling
Future Ideas
- Personas: Configure models with different perspectives (optimist, critic, pragmatist)
- Voting: Models vote on proposals, user sees tally
- Memory: Cross-project context (reference past conclusions)
- Triggers: Automated discussions on schedule or webhook
- Voice: Voice memos transcribed and fed to discussion
Created: 2026-01-16 Status: Planning Phase: 1 (Telegram POC)