A multi-backend AI operator runtime for your Mac — commanding Claude, Codex, and Gemini simultaneously.
Not a bot. Not a wrapper. A resident intelligence with 50 skills, persistent memory, and a live command center.
Website · Repository · Live Demo
The dashboard is an isometric pixel-art command centre that shows what your operator is doing in real time. Rooms light up, agents spawn, data flows pulse between stations — all driven by Server-Sent Events from the runtime.
Click any room to see the active worker, current task, model, duration, and recent history.
More screenshots
| Expandable Ticker | Koda Walking Between Rooms |
|---|---|
 |
 |
What the dashboard shows:
- Isometric rooms — Code Lab, Brain Room, Social Lounge, Memory Vault, Comms Tower, Browser Bay
- Floating room cards — click a room to see worker, task, model, duration, tool calls, history
- Animated data flows — pulses along connection paths between active rooms
- Subagent bench — idle agents (Sonnet, Haiku, Codex, Explorer, Browser) waiting to be dispatched
- Expandable ticker — scrolling recent task feed, click to expand full list
- Stats bar — active runs, completed, failed, memory promotions, learning entries, cron jobs
- Pixel-art avatars — Koda with a crown, subagents with role icons, walking animations, eye blinking, spawn bursts
Try the static demo: Open
public/demo-dashboard.htmlin your browser — no server needed.
Most AI tools start fresh every session. No memory. No context. You repeat yourself. They make the same mistakes.
Claudeking runs a 7-layer persistent memory architecture so the operator gets smarter over time:
| Layer | What it stores | Retention |
|---|---|---|
| Identity Profile | Name, rules, tone, operator config | Permanent |
| Short-Term Memory | Recent conversation context | 3 days |
| Long-Term Memory | Rules, strategies, patterns | 30 days |
| Monthly Rollup | Compacted long-term summaries | Forever |
| Semantic Embeddings | 3072-dim vectors for every conversation | Up to 20,000 entries |
| Workspace Files | Brand, logs, projects, operator config | Survives restarts |
| Background Pulse | Auto-prune, compact, pattern detection | Every 12 hours |
┌──────────────────────────────────────────────────────┐
│ You (Claude.ai Remote Control / CLI / Cowork) │
└──────────────────┬───────────────────────────────────┘
│
┌──────────────────▼───────────────────────────────────┐
│ Claude Code (native on Mac) │
│ ┌──────┐ ┌──────────┐ ┌────────┐ ┌───────────────┐ │
│ │Prompt│ │ Memory │ │ Router │ │ Command │ │
│ │Engine│ │ 7-layer │ │ Multi- │ │ Center :8787 │ │
│ │ │ │ + embed │ │ model │ │ SSE + API │ │
│ └──┬───┘ └────┬─────┘ └───┬────┘ └──────┬────────┘ │
│ │ │ │ │ │
│ ┌──▼──────────▼───────────▼──────────────▼────────┐ │
│ │ Execution Layer │ │
│ │ Code · Browser · Research · Social · Gen │ │
│ └─────────┬──────────────┬──────────────┬─────────┘ │
│ │ │ │ │
│ ┌─────────▼────┐ ┌──────▼──────┐ ┌─────▼─────────┐ │
│ │ Codex CLI │ │ Gemini CLI │ │ Memory API │ │
│ │ gpt-5.3 │ │ 3.1-pro │ │ :9090 │ │
│ └──────────────┘ └─────────────┘ └───────────────┘ │
│ │
│ ┌────────────┐ ┌──────────┐ ┌──────────────────┐ │
│ │ Telegram │ │ Notion │ │ Comet Browser │ │
│ │ Notify │ │ Logging │ │ CDP + Extension │ │
│ └────────────┘ └──────────┘ └──────────────────┘ │
└──────────────────────────────────────────────────────┘
Key design decisions:
- Multi-backend execution — Claude Code as primary, plus Codex CLI and Gemini CLI as subprocess backends. All receive the operator system prompt.
- Auto model routing — Browser/web tasks auto-route to Sonnet 4.6. Opus for quality. Haiku for routing.
- Live command center — Dashboard on
:8787with SSE, control panel, memory recall, backend status. - Telegram is notifications only — task done, cron alerts, error reports. Never a command interface.
- Session continuity — 7-day timeout. Sessions survive restarts, overnight, weekends.
- Memory-first — every prompt includes identity, session state, last 3 exchanges, and semantic recall.
git clone https://github.com/SamDev1303/claudeking.cloud.git
cd claudeking.cloud
npm install
cp .env.example .env # fill in your keys
npm run devnpm run typecheck # type-check
npm run build # compile
npm run start:resilient # productionThe main interface is Claude.ai Remote Control pointed at this workspace. Telegram handles outbound notifications.
Key env vars (see .env.example for the full list):
# Core
TELEGRAM_BOT_TOKEN= # Outbound notifications only
ALLOWED_USER_IDS= # Your Telegram user ID
BRAIN_MODEL=claude-opus-4-6 # Quality model
WORKER_MODEL=claude-sonnet-4-6 # Speed model
# Operator Identity
OPERATOR_NAME=Koda # Default operator name
OPERATOR_OWNER=Sam # Your name
# Memory & API
NOTION_API_KEY= # Execution Log, Memory Bank, Learning Ledger
MEMORY_API_ENABLED=true # HTTP API for Claude Code sessions (port 9090)
GOOGLE_AI_KEY= # Gemini CLI + embeddings (gemini-embedding-001)
# CLI Backends (auto-detected from PATH)
# Codex CLI: ~/.nvm/versions/node/v22.16.0/bin/codex
# Gemini CLI: ~/.npm-global/bin/geminiThe command centre runs as an embedded server on port 8787:
| Endpoint | Method | Purpose |
|---|---|---|
/dashboard |
GET | Serve the HTML dashboard |
/api/dashboard/stream |
GET | SSE stream for real-time updates |
/api/dashboard/state |
GET | Full state snapshot |
/api/dashboard/event |
POST | Ingest task lifecycle events |
Architecture: Hybrid Canvas + DOM. The isometric world (rooms, avatars, paths, particles, stars) renders on a canvas at 60fps with retina scaling. UI overlays (room cards, ticker, stats bar) are DOM elements positioned over the canvas using isometric-to-screen coordinate transforms.
Data flow:
Claude Code hook → POST /api/dashboard/event → dashboardStore → SSE broadcast → browser
Room cards show:
- Active worker name + avatar color
- Current task prompt (truncated)
- Model being used (color-coded)
- Live duration timer
- Last tool call (e.g.,
Read: src/index.ts) - Room history (last 3 completed tasks)
HTTP service on port 9090 for Claude Code sessions:
| Method | Route | Purpose |
|---|---|---|
| GET | /api/health |
Health check |
| GET | /api/memory/:chatId |
Read memory snapshot |
| POST | /api/memory/:chatId/record |
Record conversation turn |
| POST | /api/embeddings/search |
Semantic search across memory |
| POST | /api/embeddings/recall |
Prompt-ready recall |
| GET | /api/sessions |
List active sessions |
| GET | /api/capabilities |
Full runtime capability snapshot |
| GET | /api/skills |
Skills registry |
Auth: Bearer token if MEMORY_API_TOKEN is set, otherwise localhost-only.
Koda commands multiple AI backends simultaneously. Each receives the operator system prompt.
| Backend | Model | Binary | Use Case |
|---|---|---|---|
| Claude CLI | claude-opus-4-6 / claude-sonnet-4-6 | claude |
Primary execution, quality work, agentic tasks |
| Codex CLI | gpt-5.3-codex | codex exec |
Autonomous code writing, refactoring, builds |
| Codex Spark | gpt-5.3-codex-spark | codex exec -m gpt-5.3-codex-spark |
Near-instant responses (1000+ tok/s) |
| Gemini CLI | gemini-3.1-pro-preview | gemini -p |
Complex reasoning, 1M token context |
| Task | Model | Why |
|---|---|---|
| Final output, decisions | Claude Opus 4.6 | Owns quality |
| Browser/web/media tasks | Claude Sonnet 4.6 | Auto-routed, best at agentic web work |
| Routing, classification | Claude Haiku 4.5 | Cheapest Claude |
| Aux tasks, drafts | Gemini Flash | Fast + cheap |
| Embeddings | gemini-embedding-001 |
3072-dim, free tier (~2M tokens/month) |
- Resident Claude Code — native tool execution, always-on, operator system prompt injected
- Multi-backend execution — Claude Code + Codex CLI + Gemini CLI running simultaneously
- Remote Control — Claude.ai connects to your Mac, you watch it work
- Live Command Centre — isometric pixel-art dashboard + full control panel with SSE
- 7-layer memory — identity, short-term, long-term, rollups, embeddings, workspace, pulse
- Semantic recall — Gemini embeddings (3072-dim, ~2M tokens/month free tier)
- Auto model routing — browser/web → Sonnet 4.6, quality → Opus 4.6, routing → Haiku 4.5
- 50+ skills — auto-discovered playbooks covering code, social, media, research, deployment
- Telegram notifications — task done, cron alerts, error reports (outbound only)
- Notion logging — Execution Log, Memory Bank, Learning Ledger, Cron Jobs
- Visual generation — infographic, video, slides via browser workflows (NotebookLM, Grok, Gemini)
- Social posting — X (OAuth 1.0a) + LinkedIn (REST v2) via direct API
- Browser automation — Comet (Perplexity Chromium) via Claude-in-Chrome extension + CDP fallback
- Memory API — HTTP service on port 9090 with 16+ endpoints for sessions, recall, embeddings
claudeking.cloud/
├── src/ TypeScript source
│ ├── index.ts Entry: boot, auth, polling
│ ├── config/ Runtime config, paths, defaults, helpers
│ ├── api/ Memory API (HTTP, port 9090)
│ ├── claude/ Claude CLI client, sessions, prompts
│ ├── codex/ Codex CLI subprocess client + models
│ ├── gemini/ Gemini CLI subprocess client (NEW)
│ ├── execution/ Execution layer adapters
│ ├── telegram/ Telegram notification layer
│ ├── memory/ Store, profiles, embeddings, journals
│ ├── models/ Router, providers, aux-router
│ ├── runtime/ Dashboard server, backends, pulse, heartbeat
│ ├── integrations/ Notion, email, social, voice
│ ├── utils/ Logger, guard, env, queue, buffer
│ └── browser/ CDP connector (Comet)
├── skills/ 50+ agent skill playbooks (SKILL.md each)
├── scripts/ Standalone tools (posting, logging, worker)
├── public/ Static assets + dashboards
│ ├── command-center.html Live Koda HQ command centre
│ ├── demo-dashboard.html Static demo (no server needed)
│ └── index.html Landing page
├── claudeking/ Generated output (gitignored)
├── data/ Runtime memory (gitignored)
├── CLAUDE.md Operator config — source of truth
└── .env Secrets (gitignored)
This is an independent open-source project. Not affiliated with or endorsed by Anthropic.
Built by @kabir_Labs · claudeking.cloud