A hyper-personalized multi-agent AI assistant with a virtual pixel office
Web UI | Telegram | CLI | Daemon
JINXUS is a multi-agent AI system where a single orchestrator (JINXUS_CORE) manages 28+ specialized AI agents organized into a realistic Korean IT company structure. Each agent has a unique persona with name, personality, MBTI, and work style. The system features a real-time pixel art virtual office where you can watch your AI employees work, chat, and collaborate.
You talk to one central orchestrator. It handles everything: interpreting intent, decomposing complex tasks, delegating to specialist agents, collecting results, and delivering a unified response.
The system includes a Shared Whiteboard where agents autonomously discover new information, propose actions, and execute missions — plus a Trigger Engine and Autopilot UI for managing agent autonomy levels, routines, budgets, and inter-agent communication.
- 60x40 tile map with indoor offices, hallways, and outdoor areas (smoking area, garden, terrace, parking lot)
- 16x24 chibi sprites with 2-head proportions, team-colored uniforms, and 12 hair variations
- Camera/viewport system with drag-to-scroll and wheel-to-zoom
- BFS pathfinding with smooth interpolation movement
- 28+ POIs (coffee machine, whiteboard, printer, vending machine, benches, etc.) with state tracking
- Shared Whiteboard in the hallway — agents visit, discover new memos, and auto-generate missions
- Per-agent daily schedules based on rank and team
- Emoji activity display above character heads
- Spontaneous agent chat with 92 Korean dialogue templates (zero API cost)
- Global mute ("shut up" mode) silences idle chatter across all tabs while missions run normally
- Tool-specific animations: typing, reading, thinking, searching
- Real-time SSE state updates from backend
- 28 agents across 6 teams: Executive, Development, Platform, Product, Marketing, Biz Support
- Realistic Korean IT company structure (CTO, team leads, senior engineers, etc.)
- JINXUS_CORE orchestrates all sub-agents via LangGraph
- Dynamic hiring/firing via HR system (auto-updates playground layout)
- Agent-specific tool access control (Tool Policy Engine)
- Automatic failover and task reassignment
- Team fallback: if a specialist fails, the team lead handles it directly
- 4 mission types: QUICK (instant), STANDARD (minutes), EPIC (hours), RAID (multi-agent)
- Approval gate: review agent plans before execution (auto-approve for QUICK)
- Real-time OFFICE FEED: shows agent activity, tool calls, and reports as missions execute
- Auto work notes: mission results are automatically saved as work notes on completion
- Mission console: SSE-streamed execution log with agent conversations
- 148 tools: 19 native + 129 MCP (11 servers)
- Smart Router: auto-classifies into 4 execution paths (chat/task/background/project)
- DynamicToolExecutor with continuation support
- ToolGraph v2: BFS tool discovery (~26 microseconds per query, zero API calls)
- Artifact Store: Redis-based inter-phase data sharing
- Review Loop: automatic code review + fix cycles
- AutonomousRunner v1.8.0: up to 8 hours, 50 steps background execution
- Iteration Gate: triple-check safety (completion signal + context budget + iteration count)
- Completion Signals: agents declare
[TASK_COMPLETE],[BLOCKED],[ERROR],[CONTINUE] - Redis checkpoints for crash recovery
- Task chaining with
depends_onfor building pipelines - Telegram progress reports every 15 minutes
- Shared Whiteboard: Redis-backed shared canvas — guidelines (always-active rules) + memos (auto-discovered by agents)
- Trigger Engine: 5 trigger types (cron / event / idle / interaction / threshold) with Redis persistence
- Autonomy Dial: 4-level per-agent autonomy (Watch → Plan → Confirm → Auto) — Smashing Magazine "Shared Autonomy Controls" pattern
- Inbox: Redis-based async message queue between agents (deliver/read/mark_read)
- Relevance Gate: keyword + LLM 2-stage filter — only relevant agents respond to broadcasts
- Goal Hierarchy: company → team → agent → task goals with mission linking
- Heartbeat Protocol: periodic agent wake-up with checklist-based autonomous execution
- Atomic Checkout: Redis SETNX mission locking (409 Conflict pattern)
- Budget Enforcement: per-agent monthly API cost tracking (ok → warning → hard_stop)
- Routine Engine: cron-based recurring mission generation with concurrency policies
- WaveNoter Sync: JX_SECRETARY auto-syncs voice memos from WaveNote → Whiteboard every 10 min
- Config Revision: agent configuration versioning with rollback support
- SOUL.md: file-based agent personas (SOUL.md + RULES.md) with runtime editing
- Structured Output: Pydantic schema injection + 2-stage LLM retry for reliable JSON parsing
- 3-tier: Redis (short-term) + Qdrant (long-term vectors) + SQLite (metadata)
- LLM Memory Gate: haiku decides if memory search is needed before querying Qdrant
- Automatic reflection: generates high-level insights when importance threshold is reached
- Semantic search across long-term memory
- Time-decay pruning with configurable half-life
- 10 tabs: Office, Corporation, Projects, Memory, Logs, Tools, Notes, Settings, Autopilot, Workflow
- Autopilot Tab: Control Panel (autonomy dial + heartbeat), Triggers, Comms (inter-agent inbox), Budget dashboard, Routine manager
- Workflow Tab: React Flow-based agent execution graph visualization (BFS layout, MiniMap, Controls)
- Whiteboard Panel: overlay UI with guidelines/memos tabs, CRUD, status badges, tag filtering
- Memory File Tree: hierarchical category → result nodes with importance dots, expand/collapse
- Schema-based Settings: dynamic config form from
/status/config/schemaAPI - Single source of truth:
TEAM_CONFIGinpersonas.tsdrives all team-related UI - Real-time SSE streaming with smooth typing animation
- Team channels (Matrix/Synapse integration)
- Docker log panel (VSCode-style resizable)
| Layer | Technology |
|---|---|
| Backend | FastAPI + LangGraph, Python 3.11+ |
| Frontend | Next.js 14 + Zustand + TailwindCSS |
| Memory | Redis (short-term) + Qdrant (long-term vectors) + SQLite (metadata) |
| AI | Anthropic Claude API (Sonnet 4.6 + Haiku 4.5) |
| Tools | 19 native + 11 MCP servers (129 tools) + runtime MCP loader |
| Channels | Web UI, Telegram bot, CLI, background daemon |
| Infra | Docker Compose, PM2, volume-mounted source |
JINXUS runs as a virtual Korean IT company with 28 default employees across 6 teams. Each agent has a unique persona defined in personas.py — name, MBTI, personality, speech style, work habits, and inter-agent relationships.
Default Teams: Executive (4) · Development (6) · Platform (6) · Product (6) · Marketing (4) · Biz Support (2)
Agents are managed through a full HR lifecycle — hire, fire, and rehire at runtime without restart.
Hire (POST /hr/hire):
- Specify specialty (coding, research, writing, etc.) and the system auto-provisions everything
- Auto-assigns: persona (name, team, channel), tool policy (inherited from parent agent type), org chart placement, playground desk, team channel membership
- Appears immediately in the Pixel Office and starts receiving tasks
Fire (POST /hr/fire/{id}):
- Soft-delete: record preserved for potential rehire, agent instance removed
- Cascade: firing a team lead automatically fires all their subordinates
- Auto-cleanup: persona unregistered, removed from playground and channels
Rehire (POST /hr/rehire/{id}):
- Reactivates a previously fired agent from the fired agents list
All personas are defined as a single source of truth in personas.py (backend) and synced to the frontend via /api/agents/personas. The frontend TEAM_CONFIG in personas.ts drives all team-related UI (colors, channels, room layouts) from one central config.
backend/jinxus/
agents/ # JINXUS_CORE + sub-agents + coding/research teams
api/routers/ # FastAPI endpoints (chat, mission, logs, agents, etc.)
core/ # Orchestrator, mission executor, whiteboard, trigger engine, AAI infra
memory/ # 3-tier memory + reflection + LLM gate
tools/ # 19 native tools + DynamicToolExecutor + MCP client
hr/ # Agent hiring/firing, company channels
config/ # Settings, MCP server definitions
frontend/src/
components/
playground/ # Pixel Office engine (modularized)
engine/ # Camera, pathfinding, scheduler, social, types
sprites/ # Character (16x24), furniture (15 types), colors, icons
map/ # 60x40 map data with indoor + outdoor zones
render/ # Emoji activity display
poi/ # POI state management
tabs/ # Office, Corporation, Projects, Memory, Logs, Tools, Notes, Settings
lib/ # API client, SSE parser, personas (TEAM_CONFIG single source)
store/ # Zustand global state
- Docker and Docker Compose
- Anthropic API key
- Node.js 18+
cd backend
cp .env.example .env # Add ANTHROPIC_API_KEY
docker compose up -d # Starts FastAPI + Redis + Qdrantcd frontend
bash dev.sh # Auto-installs deps, builds, starts via pm2| Service | URL |
|---|---|
| Web UI | http://localhost:5000 |
| Backend API | http://localhost:19000 |
| Swagger Docs | http://localhost:19000/docs |
| Version | Date | Highlights |
|---|---|---|
| v4.1.1 | 2026-04-04 | AgentCallbackMixin (tool usage fix for all agents), mission log sync fix, sequential event ordering fix, internal codename → display name conversion |
| v4.1.0 | 2026-04-03 | Shared Whiteboard (AAI Phase 2), Trigger Engine (5 types), Autopilot Tab (5 sub-panels), Workflow Tab (React Flow), JX_SECRETARY WaveNoter sync, Autonomy Dial, schema-based Settings, memory file tree, 12 Geny UI improvements |
| v4.0.0 | 2026-04-01 | AAI infrastructure (Inbox, Goals, Heartbeat, Budget, Routine, Mission Lock, Config Revision), SOUL.md agent personas, Iteration Gate triple-check, LLM memory gate, Structured Output retry, Toast gating, 31 new API endpoints |
| v3.1.1 | 2026-03-26 | SSE Geny pattern, session management, agent rename, Telegram integration |
| v3.0.0 | 2026-03-24 | Pixel Office overhaul (60x40 map, 16x24 chibi sprites, camera system, outdoor areas), org restructure (realistic Korean IT company), mission real-time feed, global mute, tab rename (English), modular architecture (15 modules) |
| v2.9.0 | 2026-03-22 | Pixel Office Playground, UI overhaul, SSE agent streaming |
| v2.6.0 | 2026-03-19 | 5 new dev team agents (27 total), persona system |
| v2.5.0 | 2026-03-19 | Next.js architecture fixes, bundle optimization |
| v2.3.0 | 2026-03-16 | Smart Router, Artifact Store, Review Loop |
| v1.7.0 | 2026-03-12 | Background tasks: Redis checkpoints, pause/resume |
| v1.5.0 | 2026-03-09 | Docker multi-stage build, Tool Policy API |
Private project. Not for redistribution.
Built by jinsoo96
Special thanks to CocoRoF & SonAIengine