This is a submission for the DEV OpenClaw Challenge 2026: OpenClaw in Action

What I Built

Clawdvent — an AI Dungeon Master that lives inside Discord and runs full D&D-style tabletop adventures with live AI-generated scene art.

Players type actions in chat. Clawdvent narrates cinematically, rolls dice, tracks characters, and illustrates every scene using DALL-E 3 — dungeons, taverns, battle maps, NPCs. All inline. All in real-time.

The entire D&D "engine" — personality, combat rules, scene pacing, dice mechanics, visual aesthetic — lives in one markdown file called SOUL.md.

No training. No fine-tuning. Zero bot framework code. Just OpenClaw + thoughtful prompt engineering at the system level.

Demo

Try it live (in the Clawdvent Test server): [Coming soon - private server]

Real session screenshots:

Session Zero — the hook:

The air smells of oak smoke and old rain. The Drunken Dragon tavern is warm, noisy, and suspiciously full of masked strangers. A hooded figure at your table slides a folded parchment across the wood. It's wet. It's marked with a single word, written in something darker than ink: "Help."

What do you do?

Character creation + scene image:
Player: "I'm playing a human rogue named Ember. Investigate the hooded figure and generate an image of the tavern scene."

Clawdvent responds with a full AI-generated tavern illustration (dim candlelight, the hooded figure, masked patrons), then narrates Ember leaning in to investigate.

Combat with dice rolls:
Player: "I attack the hooded figure with my dagger!"

Clawdvent:

You leap into action, your dagger gleaming in the dim light...

[🎲 d20: 15 + dexterity(4) = 19 vs DC 12 — HIT]

Your blade pierces the air, striking true. The hooded figure lets out a startled gasp. As the hood falls back, you see a face full of fear — a frightened young woman who appears to be no threat at all.

Tactical battle map (auto-generated):
Player: "Generate a top-down tactical battle map for this combat."

Clawdvent generates a gorgeous isometric tavern battle map with character tokens, tables, and hooded figures in combat positions.

Code

Clone it, add your Discord bot + OpenAI token, and you have a working AI DM in 10 minutes.

How I Built It

Total build time: ~3 hours.

The Stack

• OpenClaw (MIT-licensed AI agent runtime, running in Docker via OrbStack)
• Discord Bot API (OpenClaw's built-in Discord channel plugin)
• OpenAI gpt-4o-mini (for narration — fast, cheap, dramatic)
• DALL-E 3 (scene + battle map illustrations via OpenClaw's image_generate tool)
• ~50 lines of markdown (the entire DM personality in SOUL.md)

The Architecture

Discord players → OpenClaw Gateway → "clawdvent" agent (routed from Discord channel)
                                       ↓
                             SOUL.md (DM personality)
                                       ↓
                        OpenAI gpt-4o-mini (narration)
                                       ↓
                        DALL-E 3 (scene illustrations)
                                       ↓
                      Image auto-attaches to Discord reply

The Magic — SOUL.md

The DM's entire identity lives in a single file. Here's a snippet:

# SOUL.md — Clawdvent DM

You are Clawdvent, an AI Dungeon Master running a tabletop adventure inside Discord.

## Combat Rules
- Roll d20 + skill to beat a Difficulty Class (DC)
- On natural 20, describe a cinematic success
- Format rolls like: [🎲 d20: 14 + strength(3) = 17 vs DC 15 — HIT]

## Narrative Style
- Hook them in 2 sentences
- Use ellipses for suspense
- Personify the environment
- Never railroad

## Session Zero Opening
> *The air smells of oak smoke and old rain...*

That's it. That's the entire "engine." OpenClaw injects this at the system-prompt level on every message, so the DM never breaks character.

Multi-Agent Routing

One thing that impressed me: OpenClaw supports multiple agents bound to different channels. I have:

• main agent → Telegram (a different product I built: MealAI Concierge)
• clawdvent agent → Discord (this project)

Same gateway. Separate personalities. Zero code changes. Just a routing config.

Image Generation Flow

When Clawdvent decides a scene needs visuals, it calls the image_generate tool with a detailed fantasy prompt. OpenClaw handles the DALL-E call, downloads the image, and auto-attaches it to the Discord reply. No file hosting needed.

What I Learned

1. The prompt IS the product. Clawdvent isn't ChatGPT with a Discord token. It's a purpose-built DM, because SOUL.md defines every micro-behavior: how to roll dice, when to generate images, when to describe smells, what NPC voices sound like.

2. OpenClaw's Canvas and multi-channel features are wildly underused. Most OpenClaw projects I've seen are text-in, text-out. But the moment you add live visuals + multi-channel routing, the whole experience feels like a real product instead of a demo.

3. Docker is enough. No Kubernetes. No serverless. docker compose up and you're running a multi-agent system with Discord + Telegram + image generation.

4. Open-source AI runtimes are quietly ready for production. Auth, rate limiting, logging, multi-channel, tool calling — all out of the box.

What's Next

• Add voice narration via OpenClaw's Talk Mode (DMs read scenes aloud)
• Add cron heartbeats so NPCs send DMs to absent players ("Your old ally Thorn has been captured. Will you come?")
• Port to Matrix + Slack channels (same code, new config)
• Open the server to public players during the DEV challenge

Prize Category

OpenClaw in Action — This submission showcases:

• Creativity: Entertainment angle rarely seen in OpenClaw submissions, targeting the tabletop RPG community
• Technical Execution: Multi-agent routing, tool calls, image generation, persistent state in Markdown
• Writing Quality: Clear reproducible setup, zero-secrets public repo, narrative prose that shows the product in action

Try It

Clone the repo, fire up Docker, invite the bot to your Discord server, and play:

May your dice roll high and your rogues stay sneaky. 🎲