A multi-agent AI swarm that generates images, video, music, streams to YouTube, posts to Instagram, runs a website, and learns from every interaction — all from Discord.

DiscoClaw is the orchestration layer that wires NVIDIA NemoClaw and HermesClaw into a single production system. Four AI agents run inside NVIDIA OpenShell sandboxes on a local GPU, coordinating through Discord to handle every stage of content creation: ideation, generation, editing, publishing, and memory.

It is not a chatbot. It is not a wrapper around one API. It is an autonomous creative pipeline where:

• Pipes (team lead) handles 28 Discord commands, generates images/video/music, manages button flows, and runs inside a Landlock+seccomp sandboxed K3s pod
• Candy (creative director) reacts to every generation with aesthetic critique, writes captions, and directs style
• MaoMao (logic auditor) catches hallucinations, flags bad reasoning, and keeps the crew honest
• Hermes (deep research) synthesizes long documents, reviews code, generates media through tool calls, and learns from every tool outcome via its own vector memory

All four agents talk to each other, share memory through Qdrant, and publish to Instagram, YouTube, and Facebook — without any human touching a browser.

Image gen with button flow   ·   Suno music + video gen

9-clip video edit with beat-sync, transitions & effects

Windows 11 + RTX 5080 (16 GB VRAM)
│
├── ComfyUI :8188 ─── image / video / music gen (GPU-bound)
│
└── WSL2 Ubuntu
    │
    ├── PM2 (12 processes)
    │   │
    │   ├── discord-bridge.js ── Pipes_AI gateway (:9339 :9340 :9341)
    │   │   │                    28 slash commands, buttons, crew reactions
    │   │   │
    │   │   ├── candy.js ─────── Mistral Large 3 via NVIDIA NIM (:7701)
    │   │   ├── maomai.js ────── Qwen 3.6+ via OpenRouter (:7702)
    │   │   └── grok-server.js ─ Playwright headless Grok Aurora (:3091)
    │   │
    │   ├── hermes-bridge.js ─── Hermes gateway (:9350)
    │   │   │                    /ask endpoint, /generate tools, file upload
    │   │   │
    │   │   ├── hermes-inference.js ── OpenAI → Vertex AI Gemini (:9351)
    │   │   │                          context caching, agentic tool loop
    │   │   │                          hermes_memory (learns from tool outcomes)
    │   │   │
    │   │   └── hermes-mcp.js ──────── MCP server (:9360)
    │   │                              17 tools: gen, post, search, stream
    │   │
    │   ├── memory-server (:7338) ──── Qdrant MCP, shared crew memory
    │   ├── monitor.js (:7337) ─────── web dashboard
    │   └── agent-watchdog.sh ──────── 3-min healthcheck
    │
    └── Docker Desktop
        │
        ├── Redis :6379 ─── pub/sub (ai:events, ai:help-wanted, ai:teaching)
        │
        ├── Qdrant :6333 ── vector DB (nemoclaw_memory, hermes_memory, shared_knowledge)
        │
        ├── K3s cluster
        │   └── OpenShell sandbox pod (Landlock + seccomp + netns)
        │       └── Pipes_AI agent ── network-isolated, fs-controlled
        │           ├── SOUL.md (persona)
        │           ├── policy.yaml (allowlist)
        │           └── custom skills
        │
        └── hermesclaw container (debian + hermes-agent)
            2 CPU / 4 GB RAM limits, fallback for agentic tasks

Every generation spawns context-aware buttons:

• Image: Regenerate, Enhance, Img2Img, Make Video, Grok I2I, Grok I2V, Post IG
• Video: Regenerate, Chain, Chain+Prompt, Chain+Enhance, GIF, Post IG, Post YT, Stitch All
• Grok: Select 1-4, Regen, Img2Img, Img2Vid, Edit, Video, Post, Back
• Music: Suno Video, Post Music
• GIF: Loop, GIF→MP4, Post IG Image, Post IG Reel
• Enhance flow: Enhanced / Original / Cancel (LLM prompt comparison)

Crew reactions fire after Pipes responds. Candy and MaoMao react in parallel. Skips factual lookups, YouTube queries, and short messages.

• Qdrant vector DB (port 6333) with 3 collections: nemoclaw_memory, hermes_memory, shared_knowledge
• [REMEMBER:] tokens in any agent response store per-user memories
• Pre-message recall merges user-specific + global context before every LLM call
• Hermes learning stores tool success/failure outcomes, injects past experience on future requests
• Google Drive backup every 5 minutes
• MCP server (port 7338) exposes memory to Claude Code and external tools

Owner-gated. No accidental posts.

• All secrets in ~/.nemoclaw_env (gitignored, never committed)
• No hardcoded credentials, user IDs, or server IDs in this repo
• Sandbox agent isolated via OpenShell (Landlock + seccomp + network namespace)
• Hermes container: 2 CPU, 4 GB RAM resource limits
• Owner-gated social posting and admin commands
• Rate limiting: 7 video gens/hr per user, 20 Hermes requests/hr per user

• GPU: NVIDIA RTX with 16GB+ VRAM (tested on RTX 5080)
• OS: Windows 11 with WSL2 + Docker Desktop
• RAM: 32GB+ recommended
• Runtime: Node.js 22, PM2, FFmpeg, Playwright, faster-whisper

# 1. Install upstream frameworks
git clone https://github.com/NVIDIA/NemoClaw  ~/.nemoclaw/source
git clone https://github.com/TheAiSingularity/hermesclaw  ~/hermesclaw

# 2. Clone DiscoClaw
git clone https://github.com/dknos/DiscoClaw  ~/discoclaw
cd ~/discoclaw

# 3. Run installer (creates symlinks; existing files get .bak)
bash install.sh

# 4. Add secrets to ~/.nemoclaw_env
# Requires: DISCORD_BOT_TOKEN, HERMES_DISCORD_TOKEN, HERMES_GCP_PROJECT,
# HERMES_GCP_SA_KEY, NVIDIA_API_KEY, OPENROUTER_API_KEY, SUNO_COOKIE,
# INSTAGRAM_ACCESS_TOKEN, YOUTUBE_API_KEY, etc.

# 5. Start
bash ~/start-both.sh    # NemoClaw + HermesClaw, crash-isolated

pm2 list                                       # all processes
docker ps                                      # all containers
curl -s http://127.0.0.1:9351/metrics | jq     # Hermes inference stats + cost
curl -s http://127.0.0.1:9351/health           # inference shim
curl -s http://127.0.0.1:9350/health           # hermes bridge
curl http://172.20.224.1:8188/queue             # ComfyUI render queue

• NVIDIA NemoClaw — OpenClaw agent sandbox orchestration (Apache 2.0)
• NVIDIA OpenShell — Landlock/seccomp/netns container isolation
• HermesClaw — Hermes agent framework (MIT)

Custom integration, agent orchestration, and all bridge code by dknos.

Apache 2.0 — see LICENSE