天守 Tenshu

天守 (tenshu) — The main tower of a Japanese castle. The highest point from which the lord surveys and commands everything below. Literally "Protector of Heaven" (天 = heaven, 守 = protector/guardian).

Real-time dashboard for OpenClaw AI agent teams. Monitor your autonomous agents with anime-styled command centers, live activity feeds, experiment tracking, and system metrics.

Screenshots

Command Center — War Room (作戦室)

AI-generated anime character portraits, per-agent task history with score trend sparklines, dust mote/firefly particles, and a Japanese dojo background.

Command Center — Control Deck (指令台)

Cyberpunk-themed mission control with neon grid, CRT scan lines, live terminal feed, agent monitor stations, and rising energy particles.

Command Center — Zen Garden (枯山水)

Peaceful cherry blossom theme with falling sakura petals, pink/stone palette, zen rock garden background, and frosted glass agent cards.

Dashboard

Live agent status cards with WebSocket connection indicator, model info, and real-time activity feed. Themed to match the active global theme.

Results Log

Experiment results from orchestrator cycles — score trends, keep/discard ratcheting, success rate stats. Based on the Karpathy autoresearch pattern.

Activity Feed

Real-time orchestrator output, per-agent logs with timestamps, and research artifact previews.

Knowledge Browser

Searchable artifact library with type filtering, stats dashboard, agent attribution, and click-to-expand content viewer.

Agent Interaction Map

Force-directed graph showing agent delegation flow with animated particles, delegation counts, and per-connection average scores.

System Monitor

Live GPU/CPU/RAM/disk metrics, loaded Ollama models, and system uptime.

Features

• 3 Immersive Themes — War Room (amber/dojo), Control Deck (cyan/cyberpunk), and Zen Garden (pink/sakura) — global toggle applies to all pages with localStorage persistence
• Command Center — Themed views with AI-generated anime character portraits, per-agent task history, score trend sparklines, live cycle status, and particle effects
• Error State UI — Red glow borders, shake animation, and error messages when agents enter error status
• Anime Sound Effects — Web Audio synthesized sounds per theme: taiko drums (War Room), synth beeps (Control Deck), wind chimes (Zen Garden) — plays on agent status changes with mute toggle
• Avatar Picker — Click agent portraits to choose from 24 AI-generated character images or upload custom avatars
• Agent Dashboard — Live status, current tasks, model info, power levels, XP bars, and level badges (Genin→Hokage) for all agents with real-time WebSocket updates
• Achievement System — 10 unlockable achievements tracked from results data (First Blood, Perfect Score, Hat Trick, On Fire, etc.) displayed on Dashboard
• Session Viewer — Active Claude Code sessions with token counts, model, duration, and cost tracking
• Cron Manager — View, toggle, and manually trigger scheduled tasks
• Results Log — Experiment results with SVG score trend line chart, ratchet floor visualization, per-agent filtering, and task type performance breakdown
• Activity Feed — Real-time orchestrator output, cycle timeline waterfall chart, agent logs, and research artifact previews
• Knowledge Browser — Searchable artifact library with type filtering (research/coder/qa), stats dashboard, and full-content viewer
• Agent Interaction Map — Force-directed graph with animated particle flow showing agent-to-agent delegation patterns, delegation counts, and avg scores
• Notification Center — Bell icon with unread badge, auto-detects high scores, low scores, agent timeouts, and orchestrator errors from live data
• Demo Mode — Simulated agents with cycling states, scores, and terminal output — auto-activates when OpenClaw isn't configured, or via ?demo=true URL param
• System Monitor — GPU temperature, utilization, VRAM, CPU, RAM, disk usage, and loaded Ollama models
• Docker Support — Single-command deployment with docker compose up

Prerequisites

• Node.js 22+ (or Docker)
• OpenClaw installed and configured (~/.openclaw/openclaw.json) — optional, demo mode works without it
• Linux with NVIDIA GPU (for system monitoring — dashboard works without it)

Quick Start

git clone https://github.com/JesseRWeigel/tenshu.git
cd tenshu
npm install
npm run dev

This starts both the API server (port 3001) and the client dev server (port 5173). Open http://localhost:5173.

Docker

docker compose up --build

Open http://localhost:3001. Without an OpenClaw config mounted, the dashboard runs in demo mode with simulated agent data. To connect real agents, uncomment the volume mounts in docker-compose.yml.

Demo Mode

If OpenClaw isn't configured, Tenshu automatically enters demo mode with 5 simulated agents cycling through working/thinking/idle/error states, fake score histories, and a live terminal feed. You can also force demo mode by adding ?demo=true to any URL.

Generating Character Art (Optional)

Tenshu includes AI-generated anime character portraits. To generate your own with different styles:

# Requires ComfyUI running locally with Flux Schnell model
pip install websocket-client requests Pillow
python scripts/generate-assets.py

Generated images are saved to client/public/assets/characters/ and client/public/assets/backgrounds/.

Configuration

Create tenshu.config.ts in the project root to customize:

export default {
  openclawDir: "~/.openclaw",  // Path to OpenClaw config
  port: 3001,                   // API server port
  clientPort: 5173,             // Client dev server port
  theme: "dark",                // dark | light | system
  accentColor: "#ff6b35",       // UI accent color
};

Environment variables:

• OPENCLAW_DIR — Override config directory (default: ~/.openclaw)
• TENSHU_PORT — Override server port (default: 3001)
• RESULTS_TSV — Override results.tsv path (default: ~/clawd/team/knowledge/results.tsv)

Architecture

tenshu/
├── shared/     # TypeScript types and constants
├── server/     # Hono API server + WebSocket + file watchers
├── client/     # Vite + React + Tailwind + Shadcn/ui
└── scripts/    # Asset generation (ComfyUI)

Data flow: Server reads openclaw.json as single source of truth → polls openclaw sessions CLI for live sessions → watches agent workspaces for file changes → reads nvidia-smi and /proc for system metrics → broadcasts updates via WebSocket → client renders in real-time.

Tech Stack

License

MIT