SkillAgentSearch skills...

fukidashi-mcp

Local manga/comic translation MCP server: offline OCR, LaMa inpainting, smart typesetting, and egui QA editor for Claude Code, Codex, Cline & OpenCode.

Install / Use

claude mcp add Yuukitake -- npx -y github:Yuukitake/fukidashi-mcp

If the server publishes to npm under a different name, use that package instead — check the repo README.

About this skill
🔌

MCP Server

Model Context Protocol server

Quality Score

84/100

Supported Platforms

Claude Code
Claude Desktop
Cline
OpenAI Codex

Our assessment of fukidashi-mcp

fukidashi-mcp scores 84/100 on our quality scale, 2768th of 4,705 Development & Engineering skills we index.

Its MCP Server is 16 KB long, well organised into 27 sections with 10 code examples: a thorough specification that gives an agent plenty to work with.

It has 10 GitHub stars, so there is little community track record yet; judge it on its content.

Substance
30/30
Structure
20/20
Description
15/15
Adoption
4/20
Freshness
15/15

Maintenance, license and trust

  • The repository was last updated about 2 months ago, so fukidashi-mcp is actively maintained.
  • It is released under the Apache-2.0 license, a permissive license that allows use, modification and commercial use with attribution.
  • Its trust signals score 97/100, with no cautions. These come from repository metadata, not a code audit — read the skill file before letting an agent act on it.

fukidashi-mcp compared with similar skills

All 4 of these similar skills score higher than fukidashi-mcp; compare them before choosing.

SkillScoreStarsUpdatedFormat
fukidashi-mcp (this skill)by Yuukitake841056d agoMCP Server
taste-skillby Leonxlnx100120.0ktodayCLAUDE.md
Agent-Reachby Panniantong100114.0ktodayCLAUDE.md
hyperframesby heygen-com10088.2ktodayCLAUDE.md
rufloby ruvnet10082.3ktodayCLAUDE.md

Frequently asked questions

How do I install fukidashi-mcp?
Run claude mcp add Yuukitake -- npx -y github:Yuukitake/fukidashi-mcp. The install tabs above show the steps for each supported agent.
Which AI agents does fukidashi-mcp work with?
It is written for Claude Code, Claude Desktop, Cline and OpenAI Codex, as a MCP Server file. Other agents that read the same format can often use it too.
Is fukidashi-mcp safe to use?
It is Apache-2.0-licensed and scores 97/100 on trust signals. Skills are instructions an agent will follow, so read the file before installing it and do not approve commands you do not understand.
Is fukidashi-mcp still maintained?
The repository was last updated about 2 months ago, so fukidashi-mcp is actively maintained.
<div align="center">

吹 Fukidashi MCP

Autonomous, 100% Local Manga Scanlation Studio for AI Agents

License: Apache-2.0 Rust: 2024 Edition Platform: Windows x64 Linux & macOS Protocol: MCP v3.2 100% Local Vision Core

Fukidashi turns OpenAI Codex, Claude Code, Gemini CLI, and Antigravity into an autonomous, professional manga localization studio running on your local machine.

Zero cloud subscriptions for vision. Zero cloud fees for cleaning. 100% private OCR, inpainting, and typesetting on consumer GPU/CPU.

<br/> <div align="center"> <a href="#-demo-pipeline"><b>Demo</b></a> • <a href="#-the-problem-vs-the-fukidashi-solution"><b>Why Fukidashi</b></a> • <a href="#-system-architecture"><b>Architecture</b></a> • <a href="#-one-prompt-codex-scanlation"><b>Codex Prompt</b></a> • <a href="#-features"><b>Features</b></a> • <a href="#-quickstart"><b>Quickstart</b></a> </div> </div> <br/>

🤖 AI Agents & Crawlers: Fukidashi is designed for autonomous tool chaining via Model Context Protocol (MCP v3.2). See the Codex One-Shot Workflow for prompt recipes.


📸 Demo Pipeline

From raw scan to publication-ready localized comic in one autonomous pipeline:

<div align="center"> <img src="assets/pipeline-preview.png" alt="Fukidashi Pipeline Showcase: Before, Clean Bubbles, and Translated" width="100%"> <p><i>1. Raw comic page ➔ 2. Neural RT-DETR detection & LaMa deep inpainting ➔ 3. Agentic translation with smart Vietnamese typesetting (Patrick Hand font).</i></p> </div>

📖 The Problem vs. The Fukidashi Solution

  • ❌ The Old Way: You manually take screenshots, feed them one by one into web chatbots, copy translated text into Photoshop, spend 4 hours clone-stamping screentones, and manually re-adjust text boxes.

  • ⚡ The Fukidashi Way: Give your AI agent a single prompt:

    "Pull the latest chapter of Dandadan and localize it into Vietnamese using Patrick Hand."

    Fukidashi auto-fetches the chapter, detects speech bubbles, deeply inpaints backgrounds with LaMa, transcribes dialogue via local OCR, translates via your agent, typesets with dynamic box-fitting, and opens the native desktop QA editor for review and CBZ export.


🏗️ System Architecture

Fukidashi orchestrates a zero-cloud local vision pipeline with your AI agent acting as the contextual translation brain:

flowchart TD
    RAW["📥 MangaDex Pull / Local Raw"] --> DET["🎯 RT-DETR (Bubble & Line Detection)"]

    %% Left Branch: Visual Cleaning
    DET -->|"Bubble Masks"| LAMA["🧹 LaMa AI Engine<br/>(Deep Inpainting)"]
    LAMA -->|"Clean Canvas"| GUI["🖥️ Native egui QA Editor<br/>(60fps Side-by-Side Review & Brush)"]

    %% Right Branch: Text Recognition, Translation & Typesetting
    DET -->|"Text Bounding Boxes"| OCR["🔍 Local Baberu / Manga-OCR<br/>(Multilingual Transcription)"]
    OCR --> CODEX["🧠 OpenAI Codex / Agent<br/>(Context-Aware Translation)"]
    LORE["🛡️ Lore & Pronouns<br/><code>fukidashi_put_lore</code>"] <--> CODEX
    CODEX --> TYPE["✍️ Smart Typesetting<br/>(Dynamic Box-Fitting & Vietnamese Fonts)"]
    TYPE -->|"Typeset Overlay"| GUI

    %% Final Export
    GUI --> OUT["📦 Export Localized CBZ / ZIP Archive"]

💬 One-Prompt Codex Scanlation

Open OpenAI Codex, Claude Code, or Antigravity, and paste:

Translate the manga chapter in "D:\manga\chapter_01" to Vietnamese using Patrick Hand font.
Follow the fukidashi-comic-translation workflow:
1. Initialize the chapter & lock character lore/pronouns via fukidashi_put_lore.
2. Clean speech bubbles with local LaMa inpainting.
3. Transcribe with local OCR and localize dialogue.
4. Typeset dynamically with full Vietnamese diacritics.
5. Launch the native egui QA editor for final approval and export to CBZ.

✨ Features

| Capability | Technical Highlight | Value | | :--- | :--- | :--- | | 📥 Automated Ingest | fukidashi_pull_chapter with MangaDex & gallery-dl | Fetch any chapter or webcomic directly from prompt | | 🎯 Bubble Detection | Neural RT-DETR model fine-tuned on comic layouts | Precise bounding box & vertical/horizontal text detection | | 🧹 Deep Inpainting | LaMa ONNX engine running on local GPU/CPU | Erases Japanese text while preserving linework & screentones | | 🔍 Multilingual OCR | Baberu Vision, Manga-OCR & PP-OCRv5/v6 | Zero-cloud transcription for Japanese, Korean, Chinese & English | | ✍️ Smart Typesetting | Dynamic box-fitting, automated font scaling | Flawless Vietnamese diacritic support (Patrick Hand & Comic Neue) | | 🖥️ Native egui Editor | Pure Rust 60fps desktop GUI (fukidashi-editor) | Zero Electron/browser bloat; inpainting brush & live bubble tweaker | | 🛡️ Strict Lore Guard | Canonical character & pronoun synchronization | Prevents pronoun drift (anh/em, cậu/tớ) across chapters |


⚡ Battle-Tested Hardware & Verified Models

Fukidashi was engineered and field-tested on real-world consumer hardware:

  • Field Test Rig: Intel Core i5-4590 (Haswell) | NVIDIA GTX 1060 6GB (Pascal) | 16GB DDR3 RAM
  • Verdict: If it runs smoothly on a 10-year-old Haswell CPU with DDR3 RAM, it will fly on your machine.

🧠 Tested LLM Translation Engines

While Fukidashi's vision core (RT-DETR, LaMa inpainting, OCR, typesetting, and desktop editor) is 100% local and requires zero external API keys, the translation brain connects to your chosen agent harness:

| Model | Type | Field Notes | | :--- | :--- | :--- | | GPT-5.6 | Cloud | 📐 High-Precision Logic. Flawless tool-calling and structured translation JSON output. | | Grok 4.6 | Cloud | 👑 Uncensored King for R18 / Doujinshi. Zero euphemisms, raw dirty talk, natural slang. | | Claude Sonnet 4.6 | Cloud | 🎯 Top-tier Literary Scanlation. Nuanced character pacing, emotional depth, fluid dialogue. | | Claude Haiku 4.5 | Cloud | ⚡ Fast, highly cost-effective daily driver for standard chapters. | | Gemini 3.8 Flash | Cloud | ⚡ Instantaneous responses, great for bulk scanning. |

🦙 100% Offline (Local LLMs)

Prefer a completely offline pipeline with zero network requests? Wire local LLMs (via Ollama, vLLM, or llama.cpp) directly into your harness:

  • Recommended Minimum: Models in the 20B+ parameter class for Japanese manga idioms and vertical reading nuances (gpt-oss:20B, gemma-4:26B).

🚀 Quickstart

🪟 Windows (Recommended - 1-Click Installer)

  1. Download Fukidashi-Setup.exe from GitHub Releases.
  2. Run the installer:
    • No Administrator rights required (PrivilegesRequired=lowest).
    • Choose any drive (D:\, E:\) to keep your C:\ drive clean.
    • Hit Next ➔ Next ➔ Finish.
  3. The installer provisions all pre-trained ONNX models (~470MB) and automatically wires Fukidashi into every supported AI client.

🐧 Linux & macOS (From Source)

# 1. Clone and build the headless server & native egui editor
git clone https://github.com/Kyokkei/fukidashi-mcp.git
cd fukidashi-mcp
cargo build --release --features editor

# 2. Auto-wire into every supported AI client
./target/release/fukidashi-mcp install

🔌 Supported AI Clients & Harnesses

Fukidashi provides native auto-wiring across all major developer and consumer AI environments:

| AI Client / Harness | Mode | Auto-Configuration Target | | :--- | :--- | :--- | | OpenAI Codex | CLI & Desktop IDE | ✅ ~/.codex/config.toml | | Google Antigravity (AGY) | AGY 1.0 & AGY 2.0 | ✅ ~/.gemini/config/mcp_config.json | | Claude Code | CLI Agent | ✅ ~/.claude.json | | Gemini CLI | CLI Agent | ✅ ~/.gemini/settings.json | | Cursor | AI Code Editor | ✅ ~/.cursor/mcp.json | | VS Code | Copilot / MCP Extension | ✅ Code/User/mcp.json | | Cline | IDE Extension / CLI | ✅ %USERPROFILE%\.cline\data\settings\cline_mcp_settings.json | | OpenCode | Terminal Agent / TUI | ✅ %USERPROFILE%\.config\opencode\opencode.json (or existing .jsonc) | | Grok Build | API / Custom Harness | ✅ Native MCP v3.2.0 protocol endpoint |


Cline and OpenCode on Windows

The installer registers Fukidashi as a local STDIO server. A standard install with no --client flag configures every supported client and creates missing user config directories/files. Use --client to limit a run, or --all as an explicit equivalent of the default. From a release directory, configure only Cline and OpenCode with:

.\fukidashi-mcp.exe install --client cline,opencode

The command merges the fukidashi entry into each existing config, creates a missing parent directory when needed, and keeps a backup under %LOCALAPPDATA%\Fukidashi\backups. A conflicting entry with the same name is left untouched and reported as an error so it cannot be silently replaced.

Cline's current shared config is %USERPROFILE%\.cline\data\settings\cline_mcp_settings.json for the IDE extension, CLI, and SDK. It uses mcpServers.fukidashi.command as a string with a separate args array. Open MCP Servers → Configure → Configure MCP Servers in the Cline panel to open the active file. Cline's MCP overview page also mentions %USERPROFILE%\.cline\mcp.json for the CLI; the canonical config page and current resolver use the shared data\settings path, which is the path Fukidashi configures. CLINE_MCP_SETTINGS_PATH and CLINE_DATA_DIR are honored when set:

{
  "mcpServers": {
    "fukidashi": {
      "command": "C:\\Users\\YOU\\AppData\\Local\\Fukidashi\\bin\\fukidashi-mcp.exe",
      "args": []
    }
  }
}

OpenCode's documented global config is %USERPROFILE%\.config\opencode\opencode.json; it also accepts opencode.jsonc. Fukidashi writes OpenCode's V1-compatible direct mcp.fukidashi entry. OpenCode's migration guide documents this form as supported by both V1 and V2, and the numeric per-server timeout is honored by the installed 1.x desktop runtime:

{
  "mcp": {
    "fukidashi": {
      "type": "local",
      "command": ["C:\\Users\\YOU\\AppData\\Local\\Fukidashi\\bin\\fukidashi-mcp.exe"],
      "timeout": 600000
    }
  }
}

The native V2 form is mcp.servers.<name> with a nested timeout object, but OpenCode 1.18.x logs that nested timeout as unsupported. The installer removes only an older Fukidashi entry under mcp.servers and preserves other servers.

The patcher reads JSON and JSONC (comments, trailing commas, and UTF-8 BOM), then writes valid JSON after a successful merge. Existing content is backed up before replacement. Fukidashi already speaks MCP over STDIO, so no HTTP port or extra environment variables are required. The execution timeout gives CPU-only OCR, cleaning, and typesetting enough time to return the structured page response; if an older client still times out, wait for the server job to finish and resume with its returned job_id rather than resubmitting the same token. Verify either client with fukidashi-mcp status, then restart the client and check that the Fukidashi tools appear. See the Cline MCP guide, OpenCode MCP guide, and OpenCode config locations for the client-side schema and precedence rules. Cline's [

Truncated for display — read the full file on GitHub.

Related Skills

View on GitHub
GitHub Stars10
CategoryDevelopment
Updated1mo ago
Forks1

Languages

Rust

Trust signals

97/100

From repository metadata: license, adoption, age and documentation. Not a code audit — see the Safety scan above for what the skill file itself contains.

1 info