Kanban 카드에서 작업을 실행하는 Voxyflow Personal AI 봇

⚠️ Alpha v0.0.1 — Very early software. It works (I use it daily), but expect rough edges. Moving fast. Contributors welcome — see below. Your personal AI bot for getting projects done. Select a card. Say "execute this." The AI reads the full context — project, description, checklist, linked files — and does the work. Not a generic chatbot. A personal agent that knows exactly where it is in your workflow. Built as a Progressive Web App. Runs locally. No cloud lock-in. Most tools solve half the problem: | Tool | What it does well | What's missing | |---|---|---| | Linear / Jira | Organizes your work | Can't execute anything | | Cursor / Copilot | Executes code tasks | Has no idea what your project is or what needs doing | | ChatGPT / Claude | Answers questions | Generic — zero project context, freezes while working | Voxyflow is the bridge. It's your kanban board and your execution engine in one — with the full context of your project always available to the AI. Think: Linear + Cursor in one app, on your own machine, with no subscription or cloud lock-in. Every card in Voxyflow is a rich context object: - Title, description, priority - Checklist items with completion tracking - Attachments and linked documents - Comments, history, relations When you click a card and say "execute this", the AI agent doesn't just see a card title — it gets the full card context, the project context, and your personality/memory files. It knows what the card is for, what's been done, and what needs doing next. This is context-scoped execution at 3 levels: General Chat → Project Chat → Card Chat (broad) (project context) (execute this exact task) At the card level, the agent has maximum context and minimum ambiguity. It doesn't ask what you mean — it executes. This is a core design principle, not a nice-to-have. Most AI tools work like this: you ask something → the app freezes → 30 seconds later, you get a response. If the task is complex, you wait 2, 5, even 10 minutes staring at a spinner. Voxyflow doesn't work that way. You ──────────────────────────────────────────────────────▶ (always talking) │ ▲ │ dispatch │ result arrives ▼ │ Worker ──── working in background (30s, 2min, 5min) ───┘ The Dispatcher (Chat Agent) handles your conversation — always responsive, zero tools, pure dialogue. When it detects a task, it spawns a Worker in the background. The Worker executes (research, CRUD, code, whatever), and when it's done, the result arrives in your conversation naturally. You never wait. You keep talking, thinking, planning — and results show up when they're ready. This is what it means to have a truly non-blocking AI assistant. ┌─────────────────────────────────┐ │ Browser (PWA) │ │ Vanilla TypeScript + Webpack │ │ HTTPS :3000 │ └────────────┬────────────────────┘ │ REST + WebSocket ┌────────────▼────────────────────┐ │ FastAPI Backend :8000 │ │ ├─ Chat Agent (Dispatcher) │ │ ├─ Workers (background exec) │ │ ├─ Analyzer (passive observer)│ │ ├─ Tool System (workers only) │ │ ├─ MCP Server (SSE + stdio) │ │ ├─ RAG (ChromaDB) │ │ ├─ APScheduler │ │ └─ SQLite (aiosqlite) │ └────────────┬────────────────────┘ │ OpenAI-compatible API ┌────────────▼────────────────────┐ │ claude-max-api proxy :3457 │ │ → Claude Sonnet / Opus │ └─────────────────────────────────┘ │ ┌────────────▼────────────────────┐ │ XTTS v2 (Corsair, GPU) │ │ → Voice synthesis │ └─────────────────────────────────┘ - Select any card on the kanban board - The chat context shifts to Card Chat — the AI has full card context - Say "execute this", "implement this", "write the tests for this card" - The Worker agent reads the entire card (title, description, checklist, attachments, history) and executes - Result streams back to your conversation without blocking anything - Chat Agent (Dispatcher) — Pure conversation. No tools. Always responsive. Dispatches work to Workers. - Workers — Background agents that execute real tasks (CRUD, research, code, file ops) without blocking the conversation. - Routed by model: Haiku (simple CRUD), Sonnet (research), Opus (complex multi-step) - Workers can run for 30 seconds, 2 minutes, 5 minutes — you keep talking the whole time - Analyzer — Passive background observer that watches conversations and auto-detects opportunities (card suggestions, patterns, action items) - Results arrive in conversation when ready — no polling, no waiting, no frozen UI - Kanban Board — Drag-and-drop columns: Idea → Todo → In Progress → Done → Archived - Roadmap — Gantt-style timeline view of cards - Sprint Planner — Group cards into time-boxed sprints - Stats Dashboard — Progress charts and velocity metrics - Wiki — Markdown documentation pages per project - Docs / RAG — Upload documents (txt, md, pdf, docx, xlsx) for AI context - GitHub Panel — Link repos, view issues and PRs - Tech Stack Detection — Auto-detect and display project technologies - Export / Import — Full project snapshots as JSON Cards are the core unit of everything: - Title, description, st