manga_mcp

An MCP server that lets Claude generate original, English-dialogue manga in a configurable Japanese art style — with a story-craft layer designed to stop output from reading like generic AI "slop."

How it avoids slop

1. Writer's pass first. manga_generate_script asks Claude (via your own Anthropic API key) to break your story prompt into a real script — capped at 10 panels total (fewer = higher quality per panel), with genre rules baked in:

   • romance_bittersweet — the core conflict must actually resolve, but at least one thread stays unresolved. No saccharine endings, no pure tragedy.

   • action_mystery — a planted clue must pay off, stakes must escalate panel-over-panel, exactly one misdirection.

2. Character consistency. The script includes one character_sheet paragraph, reused verbatim in every panel's image prompt.

3. Text is never left to the diffusion model. Dialogue/captions/SFX are drawn on top by the compositor (manga_compose_page), in real English text — not the garbled lettering diffusion models produce.

Related MCP server: OpenWebGAL Assistant

Project layout

manga_mcp/
  server.py          # MCP tool registrations (the entrypoint)
  story_engine.py     # Story prompt -> panel script (Anthropic API)
  image_backend.py    # Panel script -> panel art (Replicate API)
  compositor.py        # Panel art + dialogue -> finished page (Pillow)
  styles.py            # Art-direction presets ("our own style" lives here)
  models.py             # Shared Pydantic schemas
  config.py             # Env var loading
requirements.txt
.env.example

Setup

1. Install dependencies (Python 3.10+):

   pip install -r requirements.txt

2. Get your API keys:

   • Anthropic: https://console.anthropic.com/ → API Keys

   • Replicate: https://replicate.com/account/api-tokens

3. Copy .env.example to .env and fill in both keys.

4. Pick an art model. Browse https://replicate.com/collections/generate-anime, open a model you like (e.g. an Animagine XL / Pony-style / NoobAI-XL checkpoint), copy its owner/name:version string from the API tab, and paste it into replicate_model_version for each profile in styles.py. This is the main knob for "our own style" — tweak art_direction / negative_prompt per profile to push the look further.

5. Run it standalone to sanity-check:

   python -m manga_mcp.server

6. Connect it to Claude — two ways:

   A) Local, via Claude Desktop. Add to claude_desktop_config.json:

   {
     "mcpServers": {
       "manga_mcp": {
         "command": "python",
         "args": ["-m", "manga_mcp.server"],
         "cwd": "/absolute/path/to/manga_mcp",
         "env": {
           "ANTHROPIC_API_KEY": "sk-ant-...",
           "REPLICATE_API_TOKEN": "r8_..."
         }
       }
     }
   }

   Restart Claude Desktop and the five manga_* tools show up as available tools in chat. Requires the Desktop app to be running.

   B) Remote, no Desktop needed — works from Claude mobile too. Custom connectors connect to a server reachable over the public internet, from any Claude client (web, mobile, Desktop). Fastest path using the included render.yaml blueprint:

   1. Push this folder to a GitHub repo (git init && git add . && git commit -m init, create an empty repo on GitHub, git push).

   2. On render.com (free account): New + → Blueprint → connect that repo. Render reads render.yaml and configures the service automatically.

   3. It'll prompt you for the two secret values it can't guess: ANTHROPIC_API_KEY and REPLICATE_API_TOKEN. Paste those in.

   4. Deploy. You get a URL like https://manga-mcp-xxxx.onrender.com. Free tier cold-starts after idling — fine for testing, a $7/mo tier removes that.

   5. In Claude: Settings → Connectors → Add custom connector → paste https://manga-mcp-xxxx.onrender.com/mcp. Now usable from any chat, phone included.

Tools

Known gaps / next steps

• Character consistency is prompt-only right now. It's driven entirely by reusing the character_sheet text — good enough for a prototype, but true cross-panel consistency needs image-to-image referencing (e.g. an IP-Adapter / ControlNet reference pass, or generating a character turnaround once and feeding it back into every panel call).

• Layouts are template-based (compositor.py: LAYOUT_TEMPLATES), not yet driven by panel_size_weight — a splash panel currently doesn't visually dominate the page yet.

• No caching — re-running manga_generate_full_story re-generates every image from scratch (costs add up on Replicate).