See your AI agents work. Realm visualizes Claude Code activity in real-time 3D, orchestrates multi-agent tasks, and connects external systems via REST API.

Every tool call animates a character moving through a living workshop. Each session gets its own hexagonal zone. Tasks decompose automatically and route to the right agent. Watch it all unfold in real-time.

Visualize — Every tool call (read, write, bash, grep) triggers a context-aware animation. Reading files? Walk to the bookshelf. Running tests? The terminal glows. Committing code? Celebrate. Errors? Shake your head.

Orchestrate — Dispatch tasks to multiple Claude Code instances or Claw agents in parallel. An LLM-powered router decomposes natural language into sub-tasks and routes each to the best-fit session automatically.

Connect — External systems (OpenClaw, DingTalk, Feishu, Telegram) push tasks via REST API. Realm routes them, collects results, and fires webhooks when complete.

• macOS or Linux (Windows not supported — hooks require bash)
• Node.js 18+
• jq (brew install jq / apt install jq)
• tmux (brew install tmux / apt install tmux)

# 1. Install dependencies brew install jq tmux # 2. Configure hooks (once) npx realm setup # 3. Start npx realm

Open http://localhost:4003 and use Claude Code normally. Watch it move through the workshop as it works.

From source:

git clone <this-repo> cd Realm && npm install && npm run dev

• Multi-agent zones — Each session gets a hex zone with its own character and workstations
• Real-time 3D animations — Context-aware reactions: celebrates commits, shakes on errors
• Subagent visualization — Mini-characters spawn at the Portal for parallel sub-tasks
• Spatial audio — Synthesized sounds positioned in 3D space (Tone.js, no audio files)
• Activity feed — Live stream of tool calls, responses, and thinking
• Voice input — Speak prompts with real-time transcription (Deepgram API key required)
• Draw mode — Paint hex tiles with colors, 3D stacking, and labels
• Department grouping — Organize zones into departments (Civ-6-style)
• Station panels — Per-workstation history and details
• Attention system — Zones pulse when waiting for input or finishing
• Auto-compact / Auto-continue — Keeps long-running sessions alive
• IM integration — Receive tasks from DingTalk, Feishu, Telegram; send results back

Dispatch tasks to multiple agents and watch them work in parallel:

1. Click "+ New" (or Alt+N) to create a zone
2. Configure name, directory, agent type, and flags
3. Select a zone (1–6) to target it with prompts
4. Or use POST /dispatch to let the LLM router decide

Each session runs in its own tmux with status tracking (idle / working / offline).

See docs/ORCHESTRATION.md for the full API.

Realm accepts tasks from external systems and routes them intelligently:

IM Platform → Agent → POST /dispatch (Realm) ↓ LLM decomposes task ↓ dispatch to matching zones ↓ zones work asynchronously ↓ POST callbackUrl → Agent → IM 

Request:

POST /dispatch { "message": "...", "callbackUrl": "http://agent/callback", "sessionId": "optional" }

Callback (when all zones complete):

POST callbackUrl { "taskGroupId": "abc123", "results": [...], "durationMs": 12000 }

Configure in the Settings UI or in ~/.realm/data/settings.json:

{ "llmProviders": { "openrouter": { "provider": "custom", "apiKey": "sk-or-v1-...", "model": "stepfun/step-3.5-flash:free", "baseUrl": "https://openrouter.ai/api/v1", "maxTokens": 1024 } }, "defaultProvider": "openrouter" }

Supported: "anthropic", "openai", "custom" (OpenAI-compatible, e.g. OpenRouter, Ollama).

Draw mode: 1–6 colors · 0 eraser · Q/E brush size · R 3D stack · X clear

realm [options] realm setup # Configure Claude Code hooks realm uninstall # Remove hooks (keeps data) realm doctor # Diagnose issues Options: --port, -p <port> Server port (default: 4003) --help, -h Show help --version, -v Show version

Realm stores persistent data in ~/.realm/data/:

See docs/SETUP.md for detailed installation instructions.

See CLAUDE.md for architecture and developer reference.

MIT