Codex Integration with Creative Industry Software

What is Codex Integration with Creative Industry Software?

A set of reusable Codex skills, an MCP stdio server, and Adobe UXP/Node proxy adapters that expose safe, read‑only or dry‑run tooling for popular creative desktop software. The architecture separates Skill (high‑level AI workflow), MCP (tool registration, status, evidence), and UXP/Node Proxy (local Adobe plugin channel), with a safety layer that redacts paths and enforces sandboxed writes.

How to use Codex Integration with Creative Industry Software?

1. Clone & install

   git clone https://github.com/jianbaorui07-dot/Codex-Integration-with-Creative-Industry-Software.git
   cd Codex-Integration-with-Creative-Industry-Software
   python -m pip install --upgrade pip
   pip install -e ".[dev]"
   npm install
   

2. Run smoke checks to verify the environment.

   python -m pytest -q
   python scripts\security_check.py
   npm.cmd run preflight
   

3. Start the MCP server (Python module) and expose safe tools.

   python -m starbridge_mcp.server --json
   

4. Query tools or resources from an AI client via the MCP stdio protocol, e.g. starbridge.tools, comfyui.workflow_validate, photoshop.session_info.
5. Execute a skill using the Codex skill entry point, for example starbridge-photoshop-mcp to retrieve layer info or plan a sandboxed export.

Key Features

• Three‑layer architecture – Skill, MCP, UXP/Node Proxy – for clear separation of concerns.
• Windows‑first, local‑first design; no cloud services, all data stays on the workstation.
• Safety‑first registry – tools are read‑only or dry‑run by default; write actions require explicit confirmation and sandbox paths.
• Comprehensive bridge coverage:
  • ComfyUI workflow validation and system probing.
  • Blender environment probing, scene‑plan and reference‑reconstruction dry‑run.
  • AutoCAD/DXF plan validation, dry‑run, and guarded write.
  • Photoshop session probing, layer listing, typed BatchPlay validation (experimental write).
  • Illustrator document info and preflight (experimental write).
  • CapCut/Jianying draft directory probing.
• MCP resources & prompts – read‑only metadata, capability manifests, and reusable safety prompts.
• Extensible skill definitions – five ready‑made Codex skills for each software bridge.
• CI & pre‑flight scripts ensure that no private assets are ever committed.

Use Cases

FAQ

Q: Which operating systems are supported?

• Currently Windows only. All tooling (MCP server, UXP bridge, Node proxy) assumes Windows‑specific COM/UXP interfaces.

Q: Do I need to install the desktop applications?

• Yes. The bridge only probes existing installations of ComfyUI, Blender, AutoCAD, Photoshop, Illustrator, and CapCut/Jianying. Environment variables (e.g., BLENDER_EXE, PHOTOSHOP_EXE) tell the server where to find them.

Q: Will any private assets be uploaded to GitHub?

• No. The repository contains only code, documentation, and example scripts. All generated files, model checkpoints, PSD/AI/DWG assets, and personal credentials stay on the local machine and are excluded by .gitignore.

Q: How does the safety layer work?

• Tools are registered as safe-only by default. Calls return a dry‑run plan with evidence manifests. Write operations require a confirm_write flag and are confined to a sandbox directory defined in the MCP configuration.

Q: Can I use this from a non‑Python AI client?

• Absolutely. The MCP server communicates via stdio using the Model Context Protocol, which is language‑agnostic. Any client that can read/write to the process’ stdin/stdout can invoke the registered tools.

Q: How do I add a new creative tool?

• Implement a new MCP tool module under starbridge_mcp/, register it in the safe‑only registry, expose necessary resources/prompts, and optionally add a corresponding Codex skill under .codex/skills/.

Q: What is the difference between Skill, MCP, and UXP layers?

• Skill: High‑level AI workflow definitions, routing, and security rules.
• MCP: Stable, versioned tool API (list, call, status, evidence) with built‑in safety checks.
• UXP/Node Proxy: Experimental bridge that talks directly to Adobe desktop plugins; it does not execute arbitrary scripts and respects the same sandbox constraints.

星桥三联：StarBridge Skill + MCP + UXP

仓库总定位

这个仓库只服务三件事：Codex skill 入口、StarBridge MCP 工具协议、Adobe UXP / 本地代理桥。所有公开内容都围绕这三层收敛；历史 demo、私有素材、真实工程、本机路径、账号状态和临时输出不进入 GitHub。

一句话原则：Skill 管路线，MCP 管工具，UXP / Node Proxy 管 Adobe 本地插件通道，Safety layer 管脱敏和发布边界，专业软件继续负责真实生产。

项目状态：v0.1-alpha

Photoshop, Illustrator, Blender, and CapCut write flows are experimental or planned unless a reviewed local run proves otherwise.

Codex Skills

本仓库已经拆出 5 个 Codex skill，用来按软件结构安全调用 StarBridge MCP：

三层架构

Getting Started

Quick Install

git clone https://github.com/jianbaorui07-dot/Codex-Integration-with-Creative-Industry-Software.git
cd Codex-Integration-with-Creative-Industry-Software

python -m pip install --upgrade pip
pip install -e ".[dev]"
npm install

Smoke Checks

python -m pytest -q || python -m unittest discover -s tests
python scripts\security_check.py
npm.cmd run preflight
npm.cmd run bridge:status:safe
npm.cmd run starbridge:tools:safe

如果 PowerShell 拦截 npm.ps1，优先使用 npm.cmd。

中文阅读指南

这个仓库解决什么

它把本地创作工作站拆成多条清楚的软件桥：

一句话原则：Computer Use 管 GUI 观察和交互复现，StarBridge 管结构化安全边界，MCP 管稳定工具调用，Safety layer 管脱敏验证，专业软件管真实生产，私有资产只留本机。

主要入口

仓库区域标注

当前能力

当前仓库状态是 v0.1-alpha 工程原型。公开仓库只保存说明、协议、示例脚本、workflow、MCP stdio server、状态 manifest、测试和安全检查；模型、素材、生成图、客户文件、账号、密钥、本机安装路径和真实输出都只留在用户本机。

Photoshop、Illustrator、Blender 和 CapCut 的写入流仍是 experimental 或 planned，除非有可复查的本机运行证据。

StarBridge MCP

python -m starbridge_mcp.server --json
python -m starbridge_mcp.server tools --json --safe-only
python -m starbridge_mcp.server evidence --init --json
python -m starbridge_mcp.server evidence --validate --json
python -m starbridge_mcp.server job-status --json
python -m starbridge_mcp.mcp_server
npm.cmd run starbridge:mcp

MCP 客户端可发现首批安全工具：

starbridge.status、starbridge.probe、starbridge.tools、starbridge.evidence_init、starbridge.evidence_validate、starbridge.job_status、starbridge.recipe_list、starbridge.recipe_plan、starbridge.recipe_evidence、comfyui.system_probe、comfyui.workflow_validate、blender.environment_probe、blender.scene_plan、blender.reference_reconstruction_plan、cad_autocad.environment_probe、photoshop.session_info、ps.probe、ps.document.info、ps.layers.list、ps.batchplay.validate、illustrator.document_info、illustrator.preflight、jianying_capcut.draft_probe、jianying_capcut.draft_structure、autocad_dxf.status、autocad_dxf.validate_cad_plan、autocad_dxf.create_dxf_plan、autocad_dxf.summarize_plan、autocad_dxf.write_dxf。

starbridge.recipe_* 是跨软件高层计划入口，只返回 dry-run action plan、quality gates 和标准 EvidenceManifest 预览；真实写入仍走各 bridge 自己的确认门和 sandbox 边界。

MCP Resources（只读上下文）

除工具外，StarBridge 还按 MCP 规范暴露只读资源（“资源描述客户端该知道什么，工具描述客户端能做什么”）。客户端可通过 resources/list 与 resources/read 在不调用工具的情况下拉取静态、已脱敏的上下文：

initialize 响应同时声明 resources 能力，并在 instructions 字段返回安全优先的使用说明，帮助客户端在发起任何写入前先理解 dry-run / 确认流程。

MCP Prompts（可复用安全提示词）

StarBridge 同时暴露 MCP 的第三个标准原语 Prompts（prompts/list + prompts/get），把安全优先协议（validate-first、默认 dry-run、显式确认、sandbox-only）固化进可复用、参数化的提示模板。客户端可把它们当 slash-command 使用：

至此 StarBridge 完整暴露 MCP 三大原语：Tools（能做什么）+ Resources（该知道什么）+ Prompts（怎样安全地做）。

MCP stdio 配置示例：

{
  "mcpServers": {
    "starbridge": {
      "command": "python",
      "args": ["-m", "starbridge_mcp.mcp_server"]
    }
  }
}

Adobe UXP / Node Proxy

UXP 目前作为 Adobe 桌面软件的可审计本地插件通道，不作为通用脚本执行器。

UXP 链路默认不打开私有 PSD，不读取账号状态，不执行任意 JavaScript；写入或导出必须有显式确认和 sandbox 输出策略。

本机配置

真实安装路径不要写进 GitHub。每台电脑用环境变量或本地 .env 管理：

发布边界

不要提交以下内容：

• 历史网页 demo、虚拟宠物 demo、PPT 工作区和无关临时输出。
• 报告生成临时产物、图片素材、样式参考文档。
• 输出目录、缓存目录、日志和临时文件。
• 模型文件、LoRA、VAE、ControlNet、生成图片、ComfyUI 输出目录。
• 私有 .blend、贴图、资产库、渲染缓存或商业模型。
• 客户 DWG、商业图纸、授权文件或真实 CAD 输出。
• PSD / AI 私有工程、商业字体、商业笔刷、购买素材、源图和导出结果。
• 剪映 / CapCut 草稿、缓存、导出视频、字幕原稿、账号和会员信息。
• 密码、token、Cookie、OAuth 缓存、浏览器资料和支付信息。

发布前验证

python scripts\starbridge_preflight.py --markdown
python scripts\starbridge_preflight.py --write-report --soft-exit
npm.cmd test
npm.cmd run preflight
npm.cmd run bridge:status:safe
npm.cmd run starbridge:tools:safe
python scripts\security_check.py
python scripts\bridge_capability_matrix.py --check

CI 候选检查使用跨平台命令：

python scripts/security_check.py
python scripts/collect_bridge_status.py --json
python examples/bridge_status.py --json --redact-paths --soft-exit
python -m starbridge_mcp.server tools --json --safe-only
python -m starbridge_mcp.server evidence --init --json
python -m starbridge_mcp.server evidence --validate --json
python -m starbridge_mcp.server job-status --json

Release Readiness

• Visual demo: docs/visual-demo.md
• Install and publish path: docs/install-and-publish.md
• Visual evidence: docs/adobe-demo-gallery.md
• Local smoke test: docs/adobe-demo-smoke-test.md
• Draft release notes: RELEASE_NOTES_DRAFT.md
• Capability matrix: docs/CAPABILITY_MATRIX.md
• Client compatibility: docs/client-compatibility.md
• Windows install notes: docs/windows-install.md
• Adobe AI integration map: docs/adobe-ai-agent-integration-map.md
• Illustrator vector rebuild pipeline: docs/illustrator-vector-line-rebuild-pipeline.md

下一步

完整路线图见 ROADMAP.md。

For English Readers

StarBridge is a Windows-first, local-first Skill + MCP + UXP bridge for connecting AI coding agents to creative desktop software: ComfyUI, Blender, AutoCAD / DXF, Photoshop, Illustrator, and CapCut / Jianying. It focuses on reusable Codex skills, safe MCP probes, workflow validation, redacted status reports, and guarded local automation examples instead of uploading private assets or replacing the creative tools.

Start with this README and docs/local-mcp-setup.md. Most project notes are Chinese-first because the current workstation and software setup are Windows-first, but commands, tool names, environment variables, and MCP APIs are kept in English.

Search keywords: MCP, Model Context Protocol, Codex, AI agent, creative software automation, ComfyUI workflow, Blender automation, AutoCAD DXF, Photoshop UXP, Photoshop COM, Illustrator scripting, CapCut Jianying, local-first AI tools.

License

This project is licensed under the MIT License. See LICENSE.

Copyright (c) 2025 曹瑞 and contributors.