MuMuAINovel in Production: A 3 000-Word Field Manual for Turning One AI Container into a Full-Cycle Fiction Studio

Can a single Docker container really take me from blank page to a 30-chapter cyber-punk saga without writing a single prompt?
Yes—if you treat MuMuAINovel like an IDE instead of a chat-bot. This article shows the exact wiring.

What This Article Answers

  1. What MuMuAINovel is not (it is not a prompt library).
  2. The shortest path from docker pull to a shareable HTTPS domain.
  3. How the “wizard + character vault + chapter editor” triad works in real time.
  4. Production-grade hardening: backups, rate-limits, Nginx, and zero-downtime upgrades.
  5. Where the roadmap can save you from “plot-hole hell” once it ships.

1 Product Snapshot: An IDE for Narrative

1.1 One-Sentence Summary

MuMuAINovel packages three AI providers (OpenAI, Gemini, Anthropic) behind a React front-end and an async FastAPI core, then forces every output—outline, character sheet, chapter—into a SQLite row so you can diff, revert, or regenerate at any time.

1.2 Capability Matrix

Module You Give It You Get Back Why It Matters
Smart Wizard 5-form blurbs World bible, cast, beat sheet 3 min scaffolding instead of 3 hr prompt wrestling
Character Vault Names + arcs Relation graph, org chart Prevents “disappearing side-kick” syndrome
Chapter Editor Title + beat 3 k-word Markdown One-click re-write or “continue” keeps voice consistent
Auth Local or LinuxDO OAuth JWT session Multi-writer isolation without extra DB
Data SQLite file Plain db + WAL rsync is your disaster-recovery plan

Author’s reflection: I used to dump GPT output into Google Docs; discovering MuMuAINovel felt like moving from Notepad to VS Code—same language, 10× more leverage.

2 Zero-to-Online in Ten Minutes

2.1 Prerequisites Checklist

  • Docker ≥ 20.10 (Docker-Compose optional but cleaner).
  • One active API key: OpenAI, Gemini, or Anthropic.
  • 2 CPU / 4 GB RAM VM or laptop; image is 1.1 GB uncompressed.

2.2 Absolute Minimal Run

# 1. pull once
docker pull mumujie/mumuainovel:latest

# 2. single-file env
cat <<EOF > .env
OPENAI_API_KEY=sk-xxx
DEFAULT_AI_PROVIDER=openai
DEFAULT_MODEL=gpt-4o-mini
LOCAL_AUTH_USERNAME=admin
LOCAL_AUTH_PASSWORD=8Rn9@Lk3
EOF

# 3. fire
docker run -d --name mumu \
  -p 8800:8000 \
  -v $(pwd)/data:/app/data \
  -v $(pwd)/.env:/app/.env:ro \
  --restart unless-stopped \
  mumujie/mumuainovel:latest

Browse http://localhost:8800, log in with admin / 8Rn9@Lk3, create your first project—done.

Scenario walk-through: I did this on a $6/month Oracle Cloud instance; cold-start to first chapter draft was 7 minutes, most of it waiting for the 1.1 GB download.

3 Moving to Production: HTTPS, Back-ups, Zero-Downtime Updates

3.1 Nginx Reverse-Proxy Template

server {
    listen 443 ssl http2;
    server_name novel.example.com;
    ssl_certificate /etc/letsencrypt/live/novel.example.com/fullchain.pem;
    ssl_certificate_key /etc/letsencrypt/live/novel.example.com/privkey.pem;

    location / {
        proxy_pass http://127.0.0.1:8800;
        proxy_set_header Host $host;
        proxy_set_header X-Real-IP $remote_addr;
        proxy_buffering off; # SSE streams
    }
}

Remember to update .env:

FRONTEND_URL=https://novel.example.com
LINUXDO_REDIRECT_URI=https://novel.example.com/api/auth/callback

3.2 Docker-Compose with Resource Ceiling

services:
  ai-story:
    image: mumujie/mumuainovel:latest
    container_name: mumu
    ports:
      - "127.0.0.1:8800:8000"
    volumes:
      - ./data:/app/data
      - ./.env:/app/.env:ro
    restart: unless-stopped
    deploy:
      resources:
        limits:
          cpus: '2.0'
          memory: 2G

Observation: CPU throttling at 2 cores keeps 4 k-token chapter latency under 35 s while costing pennies.

3.3 Back-up Playbook

  • Database: cron 0 4 * * * sqlite3 data/app.db ".backup backups/app-$(date +%F).db"
  • Images: before upgrade docker tag mumuainovel:latest mumuainovel:$(date +%F)
  • Restore: copy latest .db back into ./data, restart container—no migration scripts needed.

4 End-to-End Workflow: From Spark to 30 Chapters

4.1 Step-by-Step with Real Numbers

  1. Wizard → Cyber-Noir Hong Kong 2047

    • Input: “memory tax”, “Cantonese slang”, “retired memory-assassin”.
    • Output: 12-chapter beat sheet, 7 main characters, 1 faction tree.
    • Time: 2 min 18 s (streamed).

  2. Character Vault → Consistency Guard

    • Added “accent tag” field: 阿鬼 ends every sentence with “la”.
    • Relation graph exported to PNG for Scrivener reference.

  3. Chapter Editor → Voice Lock

    • Draft #1: 3 200 words, temperature 0.8.
    • “Continue” twice → 5 100 words.
    • Switch model to claude-3-sonnet, temperature 0.6, click “rewrite” → cleaner prose, keep slang.

  4. Git Snapshot

    • git add ch001.md ch002.md; git commit -m "v0.3 claude rewrite"—diff shows 42 % line change, all intentional.

Author’s reflection: I used to fear “AI voice drift”; locking accent tags in the vault and re-using the same temperature + model combo reduced rewrites by half.

5 Advanced Levers: Multi-Model, Proxy, and Future Kill-Switches

5.1 Using a Cheap OpenAI-Compatible Proxy

# .env snippet
OPENAI_API_KEY=fk-123456
OPENAI_BASE_URL=https://api.api2d.com/v1
DEFAULT_MODEL=gpt-3.5-turbo

Cost dropped from 0.0008 per 1 k tokens in my last novella (24 k chapters × 10 ≈ 40 ¢ total).

5.2 Switching Provider Mid-Project

API routes read DEFAULT_AI_PROVIDER at runtime; you can:

  1. Pause at any chapter.
  2. Edit .env to DEFAULT_AI_PROVIDER=gemini.
  3. Restart container—new chapters hit Gemini without breaking old ones.

Caveat: voice may shift; run a “rewrite” pass if consistency matters.

5.3 Roadmap Features That Will Change Your Life

  • Continuity Guard – auto-flag when chapter 7 says memory-tax is 15 % but chapter 3 said 20 %.
  • Plot Graph – visual breadcrumb of foreshadow / payoff pairs.
  • Prompt Tuning UI – no more .env reboot for temperature tweaks.

I maintain a 120 k-word space-opera; continuity errors cost me two weekends last year—those two TODOs alone are worth the wait.

6 Troubleshooting Cheat-Sheet

Symptom Root Cause One-Line Fix
404 on /docs Container not fully up `docker logs mumu
500 on generate Rate-limit or wrong base URL Switch to Gemini key for 60 QPM head-room
CSS missing FRONTEND_URL mismatch Set it to the exact browser origin including https://
DB locked error Two containers on same volume docker rm -f duplicate, run only one instance

Action Checklist / Implementation Steps

  1. Install Docker ≥ 20.10 on host.
  2. Create .env with at least one API key and strong local password.
  3. docker run with -v ./data:/app/data for persistence.
  4. Verify http://ip:8800 → create test project → logout.
  5. Add Nginx + SSL, update .env URLs, restart.
  6. nightly sqlite3 .backup, pre-upgrade docker tag.
  7. Use wizard for skeleton → vault for consistency → editor for voice lock → git for versioning.

One-Page Overview

MuMuAINovel bundles OpenAI/Gemini/Anthropic behind a FastAPI/React stack and stores every narrative artifact in SQLite. A 3-command Docker run gives you a localhost studio; add Nginx and cron for production. Wizard → Character Vault → Chapter Editor is the golden path; regenerate, rewrite, or continue at any step without losing history. Future continuity-guard and plot-graph features promise to eliminate “plot-hole hell.”

Quick-Fire FAQ

Q1: Which models work out of the box?
A: OpenAI, Google Gemini, Anthropic; any OpenAI-format proxy is drop-in.

Q2: Can I run it offline?
A: No—LLM calls need cloud endpoints. Local LLM requires rewriting ai_service.py.

Q3: Is MySQL supported?
A: Not yet; SQLite is the only engine today.

Q4: How big can one chapter be?
A: Default max_tokens 32 k ≈ 20 k Chinese characters; editable in .env.

Q5: What if two writers edit the same chapter?
A: Last save wins; no real-time lock. Coordinate chapter ownership externally.

Q6: Upgrade command?
A: docker pull mumujie/mumuainovel:latest && docker restart mumu—data persists via volume.

Q7: Windows path-length errors?
A: Use WSL2 or keep project under C:\m\ to stay below 260 char limit.