---

Source

• GitHub (primary repository): https://github.com/Justin-sky/ai-art-engine
• Gitee (CN mirror): https://gitee.com/beijing_blue_whale_era_zhangjian/ai-art-engine

The main branches of both repositories stay in sync; Releases and the client's auto-update are still anchored to GitHub.

---

Download

Platform	Package	Get it
Windows	.exe	GitHub Releases
macOS	.dmg (x64 / arm64)	GitHub Releases: choose x64 for Intel, arm64 for Apple Silicon (with arm64 only, Intel Macs will show as unsupported)
Linux	.AppImage	GitHub Releases (run chmod +x first)

Pushing a v* tag triggers GitHub Actions to automatically build and publish multi-platform installers. You can also package it yourself:

npm run dist:win    # Windows
npm run dist:mac    # macOS
npm run dist:linux  # Linux

---

Features

AIArtEngine — a pro AI tool for short dramas, ads & finished videos: assets, storyboards & node graphs in one desktop app, local-first, your own API key.

Highlight: AI Chat + MCP

The in-app chat can drive all generation and orchestration capabilities; external AI agents like Claude Code / Codex can also connect directly to operate on your projects.

• AI Chat Panel — In-app AI assistant (DeepSeek Harness runtime): @-reference project assets in the chat, and have the Agent call MCP tools to get work done directly — generate images / videos / 3D / voice, edit node graphs, run workflows and check status; multi-session history, any configured text model can be selected
• MCP Tool Service — Built-in MCP Server; external AI agents like Claude Code / Codex can connect via stdio bridge or HTTP to directly operate on your project (plan and save workflows to disk, run generation, read/write assets and node graphs); tokens are persistently reused across restarts, operations are audited, concurrency is gated

Highlight: Skill System

Skills are "job handbooks" — they tell the Agent how to do things (workflow, conventions, output format); MCP tools are "hands and feet" — they actually execute. Together, the chat can "walk the talk".

• Built-in creative skills — Automatically available in chat, no configuration needed: storyboards / animation (9-grid storyboard table, beat breakdown table, dynamic prompt table, 4-grid dynamic storyboard table, etc.), director review, system creation (script, image-to-prompt, image / video generation, sound, emotion, lighting, multi-angle, outpainting, inpainting, HD upscale, UI, UI splitting, world extraction, beat splitting, beat unit generation, pitch plan, portrait stickers, prompt optimization, eraser)
• Custom skills — Settings → Custom Skills: place .md files conforming to the dsh SKILL.md format (frontmatter name / description + Markdown body) into the skill directory, and they'll be automatically loaded in the next chat session; one-click "Generate sample template" within the directory
• Mechanism — The skill list is injected into the chat context via <available_skills>; when the Agent judges a task match, it loads the skill file on-demand as instructions; built-in skills are automatically managed by the program (snapshot + fingerprint dedup), custom skills are unaffected. Node graph nodes also use the same set of skills to define node behavior

Highlight: Blender MCP Server (Let AI Directly Operate Blender)

As long as Blender is running locally, the AI can build scenes, write materials, take screenshots for self-inspection, and export GLB in chat — outbound direct connection to the addon, no uv / Python / subprocess required.

• Directly drive Blender — 9 MCP tools covering the "see current state → write bpy modeling → screenshot self-check → export GLB" loop: get_scene_info / get_world_state_snapshot / get_object_info to query scenes and selected objects, execute_blender_code to execute Python inside the Blender process (full access to bpy / bmesh / mathutils), get_viewport_screenshot to capture the viewport view and return it to multimodal clients, export_scene to output GLB / GLTF / FBX / OBJ / USD / STL (with asset_import for one-click asset library import), describe_node_type / bpy_api_lookup / get_addon_status to query node ports, bpy API, and addon versions
• Dual addon compatibility — Supports both the community solution blender-mcp addon.py and the official solution Blender Lab "MCP Server" extension, switchable in "Settings → MCP → Blender Toolset"; under both backends, tool names, parameters, and output structures are identical, transparent to the model side
• Zero extra dependencies — No uv install, no Python subprocess, no Blender config changes: the application actively connects outbound to the addon's localhost:9876 listening port; out-of-the-box in installer / repo / container environments
• In-app chat + external Agents share the same source — The ◈ AI chat panel on the left of the workspace (say "use Blender to assemble these Cubes into a base, take a screenshot when done") and external Agents like Claude Code / Codex use the same Blender toolset; disabling in Settings removes the entire group from the tool list, without affecting the main toolset; the panel's Ask / Plan modes also constrain Blender tools — it's not a bypass for panel modes
• Safety guardrails — execute_blender_code has lexical guardrails (rejects os / subprocess / network modules, open / __import__ escape chains, structural writes like decorators / class / type() class creation; bpy operators for rendering / saving / import-export are unrestricted); screenshots use one-time temporary files, the view only returns to multimodal clients via MCP responses, not saved to the project or audit log

For detailed tool lists, protocol details, and security design, see [MCP Integration Guide](https://github.com/Justin-sky/ai-art-engine/blob/HEAD/docs/MCP.md#⑤- blender-工具集可选需要本机-blender) / MCP Integration Tutorial.

Full Capabilities

• Local projects — New / Open / recent list; JSON + media directory on disk; data never leaves your machine
• Asset library — Images / videos / audio / 3D models; AssetRef GUID; .aipackage import/export; multi-select files / multi-select folders drag to another folder (disk move + descendant asset.relativePath write-back + cycle detection + same-name append " 2" / " 3")
• Storyboard & Canvas — Shot parameters, Fabric composition, dockable layouts
• One-click workflows — Preset templates (short drama storyboard, game UI, game UA, product ad, e-commerce livestream, game 3D assets, comic publishing, knowledge talking head, 3D white-model preview, etc.) or AI-planned topology; one-click creates reusable host assets (boundary I/O + Dive inner graph)
• Node graph generation — Text / image / video / audio / 3D model nodes; instruction panel and model parameters; generation lock, asset library dual output; port types must match (singular can't enter plural, selection node only accepts list port); wire styles / minimap; task queue reuses common upstreams, task fault-tolerance mode (node failure degrades without breaking the whole chain); comic pages (storyboard panels + speech bubbles, export transparent PNG), ad variant matrix, media QC / rework (dedicated QC model five-dimension scoring → FAIL auto-injects reason and retries), 2D frame animation and frame animation sequence, layer separation (export PSD after splitting)
• Host assets — Outer boundary ports exposed, inner graph supports Dive; multiple convergence points each get independent outputs
• Director stage — 3D positioning screenshots and action recording (written to Cache/Videos); square ports out-shots / out-actions; 3D model input port, Dive auto-instantiates to stage; panorama input auto-set as background; AI white-model scene building, 20+ basic geometry types, material texture mapping (base / normal map), shading modes and wireframe mode
• Final video timeline — Asset grouping and track sequencing; picture-in-picture overlay (position / size / opacity / volume) and video track transitions; preview selected / full timeline play; export final video
• Multi-model providers — OpenRouter, OpenAI (GPT text / gpt-image images), DeepSeek (text), Zhipu (GLM text / CogView images), Kimi / Moonshot (text), xAI / Grok (text / images / videos), Google / Gemini (text / images / videos), local vLLM (text / Wan videos), Ollama / LM Studio (text, OpenAI-compatible, no API key needed), Volcano Ark (Seedream / Seedance / voice), Keling, MiniMax, Tongyi Qianwen (DashScope), ModelScope, ComfyUI (API 2: images / videos / audio, local or cloud Base URL), MagicRouter (aggregation gateway: text / images / videos), Meshy / Tripo / Rodin (Hyper3D) / Luma AI / Lux3D (3D model generation, text-to-3D / image-to-3D), custom provider (choose OpenAI-compatible / Anthropic / Gemini endpoint type, fill Base URL and API key to pull text models)
• 3D model rigging & skeleton — Check "Generate skeleton" on a 3D node for upstream skeleton rigging: Tripo sends rig=true, Meshy sends rigging=true + rig_type (humanoid etc.) + pose, Rodin (Hyper3D) sends rig=true + rig_type + rig_animation (multipart, some tiers need rig=true for rigged GLB); Luma AI / Lux3D do not support rigging, force-enabling will be rejected with a clear message
• Web search providers (configured in Settings panel) — DeepSeek Search (default) / Tavily / Brave / SerpAPI / Mock, five adapters available in parallel; each has independent toggle, Base URL, API key, search depth and time window; panel includes built-in connectivity test; deliberately separated from model provider system (fewer in number, no context injection)
• HTTP network error diagnostics generalized — axios network errors (err.code + err.cause.code / err.cause.message) extracted from the Keling adapter into a generic readHttpError, covering all 33 providers — DNS / TCP / TLS handshake pre-socket disconnect scenarios; the UI now shows stable labels like ECONNRESET / UND_ERR_SOCKET for quick firewall vs connection-refused diagnosis
• Object storage — Volcano Engine TOS, Alibaba Cloud OSS, Tencent Cloud COS (only one active at a time, used for public external links like reference videos)
• Extensible — Editor Kernel + Cordis internal extensions (windows / Inspector / nodes / Skills / executor) + declarative external plugin manifest

Models & Object Storage Overview

Type	Provider	Capabilities
Model	OpenRouter	Text / Images / Videos (aggregated catalog)
Model	OpenAI	Text / Images (requires network access and account for api.openai.com)
Model	DeepSeek	Text (deepseek-chat / deepseek-reasoner)
Model	Zhipu	GLM Text / CogView text-to-image
Model	Kimi (Moonshot)	Text (kimi-k2 series / moonshot-v1 series)
Model	xAI (Grok)	Text / Grok Imagine Images / Grok Imagine Video (async polling)
Model	Google (Gemini)	Text / Nano Banana Images / Veo 3.1 Video (async polling, official OpenAI-compatible layer)
Model	vLLM	Local Text / Video (Wan T2V / I2V, OpenAI-compatible, no API key)
Model	Ollama / LM Studio	Local Text (OpenAI-compatible, no API key)
Model	Volcano Ark	Text / Seedream Images / Seedance Video / Voice Design
Model	Keling	Images / Videos (API Key)
Model	MiniMax	Text / Images / Videos / Voice Design
Model	Tongyi Qianwen	Text (compat mode) / Wanxiang Images / Wanxiang Video (incl. HappyHorse etc.)
Model	ModelScope	Text / Text-to-Image (access token)
Model	ComfyUI	Images / Videos / Audio (API 2; local 8189 or cloud Base URL)
Model	MagicRouter	Text / Images / Videos (OpenAI-compatible aggregation gateway, async video polling)
Model	Meshy	Text-to-3D / Image-to-3D (incl. multi-image-to-3D, API Key)
Model	Tripo	Text-to-3D / Image-to-3D (API Key)
Model	Rodin (Hyper3D)	Text-to-3D / Image-to-3D (API Key)
Model	Luma AI	Text-to-3D / Image-to-3D (API Key)
Model	Custom	Text (choose endpoint type: OpenAI-compatible / Anthropic / Gemini, fill Base URL and API Key)
Model	Lux3D	Text-to-3D / Image-to-3D (incl. multi-image-to-3D, G1 / G1-Turbo, API Key)
Object Storage	Volcano TOS / Alibaba Cloud OSS / Tencent Cloud COS	Reference media upload and signed URLs; mutually exclusive in Settings

Configuration entry points: Settings → Models / Settings → Object Storage. Local ComfyUI requires installing comfy-api-proxy first (default 8189); do not connect directly to 8188.

---

Quick Start

Using the installer

1. Download the installer for your platform from Releases
2. Install & launch → New Project
3. Add a model provider and fill in your API key in Settings; optionally configure object storage
4. Use the top bar "One-Click Workflow" to quickly create a host asset, or manually build a pipeline in Storyboard / Node Graph
5. Click "◈" in the left sidebar to open AI Chat, @-reference assets to let the assistant generate directly; or follow the MCP Integration Tutorial to connect an external agent like Claude Code

For full usage instructions, see the User Manual (source in website/manual.html). Local ComfyUI requires installing comfy-api-proxy first (default port 8189); do not connect directly to 8188.

From source

# GitHub
git clone https://github.com/Justin-sky/ai-art-engine.git

# 或使用 Gitee 国内镜像
git clone https://gitee.com/beijing_blue_whale_era_zhangjian/ai-art-engine.git

cd ai-art-engine
npm install
npm run dev

Dev builds require Node.js 22+ (the runtime already bundles Node, no user install needed). If the Electron download fails, you can:

set ELECTRON_MIRROR=https://cdn.npmmirror.com/binaries/electron/
node node_modules/electron/install.js

npm run typecheck && npm test   # 检查
npm run pack                    # 未封装目录，便于自测

---

Docs

• User Manual (source website/manual.html)
• Architecture · Node Graph Plugins · Docs Index
• Asset Model · AssetRef · Asset Package
• Roadmap · Changelog

---

Versioning & Updates

• The version number is authoritative in package.json's version field (SemVer); see CHANGELOG.md for changes.
• Releasing: update package.json and CHANGELOG first, then commit, tag, and push, e.g.:

git tag v5.0.0
git push origin v5.0.0

CI verifies the tag (stripping the v prefix) matches package.json, then builds and publishes a GitHub Release (including update metadata such as latest.yml).

• Client update: after launching the installer, it checks Releases; you can also manually check under Settings → General → About & Update, then restart after download completes. Dev mode (npm run dev) does not check for updates.

---

Contribute

Issues and Pull Requests are welcome. Before submitting, please:

npm run typecheck && npm test

For discussion and bug tracking, use GitHub Issues or Gitee Issues.

---

Community

• Website: GitHub Pages · Alibaba Cloud OSS (see npm run site:deploy)

• Video tutorials: Bilibili Space

• CN mirror: Alibaba Cloud OSS static hosting (custom domain / publicBaseUrl)

• GitHub: Justin-sky/ai-art-engine

• Gitee: beijing_blue_whale_era_zhangjian/ai-art-engine

• QQ Group: 647306826 (scan to join)

  

• Email: 284139554@qq.com

---

Stack

Electron · Vue 3 · Pinia · TypeScript · Cordis · Fabric.js · electron-vite · electron-builder

---

License

GPL-3.0 — when modifying and redistributing, derivative works must be open-sourced under the same license.

---