mindcode-claw-image2

中文版 | English

A publishable draft of a thin skill + local service package that lets your agent, whether OpenClaw, 爱马仕, or another claw-style agent, use your existing GPT subscription entitlement and quota to generate images through Image 2.

Purpose

The point of this project is simple:

• let your agent use your own GPT subscription entitlement and quota
• route image generation through Image 2
• make that capability directly callable from claw-style agents such as OpenClaw, 爱马仕, and similar agent setups

This package does not create a new image model account. It exposes a stable local path so your agent can use the image-generation capability you already have.

What this package is

This repo combines two layers:

1. A thin skill (SKILL.md)

   • routes requests to the Codex imagegen path when the user explicitly asks for it
   • defaults to raw prompt pass-through
   • prevents silent fallback to HTML layout, manual infographic workflows, or unrelated image tools

2. A local service (scripts/)

   • wraps codex exec image generation behind a small HTTP API
   • provides health checks, async jobs, artifact persistence, install flow, and smoke testing

The service is the execution layer. The skill is the boundary layer.

Repo layout

mindcode-claw-image2/
├── SKILL.md
├── README.md
├── PUBLISHING.md
├── LICENSE
├── launchd/
│   └── com.openclaw.codex-imagegen-service.plist
└── scripts/
    ├── codex-imagegen-service.js
    ├── codex-imagegen-worker.js
    ├── codex-imagegen-smoke-test.js
    ├── codex-imagegenctl.js
    └── install-macos-launchd.sh

Prerequisites

• codex CLI installed locally on the user's machine
• a valid Codex or GPT subscription/path that actually permits image generation
• enough upstream quota, credits, or allowance for image generation jobs
• Node.js 18+
• macOS if you want to use the included launchd install path
• a machine where generated Codex images land under ~/.codex/generated_images

Important: this package does not provide image-generation entitlement by itself. It only wraps an already-working local Codex imagegen capability.

Install on macOS with launchd

bash scripts/install-macos-launchd.sh

What this does:

• resolves the repo root
• fills the launchd plist template
• installs it to ~/Library/LaunchAgents/
• bootstraps and kickstarts the service
• runs a basic /health check

What this does not prove:

• that the local codex account has imagegen permission
• that the upstream subscription is active
• that the account still has remaining quota or credits

Control commands

node scripts/codex-imagegenctl.js install
node scripts/codex-imagegenctl.js start
node scripts/codex-imagegenctl.js stop
node scripts/codex-imagegenctl.js restart
node scripts/codex-imagegenctl.js status
node scripts/codex-imagegenctl.js health
node scripts/codex-imagegenctl.js smoke
node scripts/codex-imagegenctl.js submit --prompt "a red cube on white background"
node scripts/codex-imagegenctl.js job <job-id>
node scripts/codex-imagegenctl.js logs --follow

Service contract

Endpoints

• GET /health
• POST /v1/images/generations
• GET /v1/jobs/:id

Behavior

• the service runs one queued job at a time
• artifacts are written to <CODEX_IMAGEGEN_ARTIFACT_ROOT>/<job-id>.png
• completed and failed jobs expire from in-memory state after CODEX_IMAGEGEN_JOB_TTL_MS
• persisted state is stored in CODEX_IMAGEGEN_JOB_STATE_FILE
• if the service restarts while a job is queued or running, that job is recovered as failed with service-restarted

Run locally without launchd

node scripts/codex-imagegen-service.js

Health check:

curl -sS http://127.0.0.1:4312/health

Submit a job:

curl -sS -X POST http://127.0.0.1:4312/v1/images/generations \
  -H 'content-type: application/json' \
  -d '{"prompt":"a minimal white image with a tiny black dot centered","timeout_sec":180}'

Smoke test:

node scripts/codex-imagegen-smoke-test.js

Configuration

Supported environment variables:

• CODEX_IMAGEGEN_PORT
• CODEX_IMAGEGEN_ARTIFACT_ROOT
• CODEX_IMAGEGEN_JOB_TTL_MS
• CODEX_IMAGEGEN_JOB_STATE_FILE
• CODEX_IMAGEGEN_WORKDIR
• CODEX_IMAGEGEN_BASE_URL
• CODEX_IMAGEGEN_SMOKE_PROMPT
• CODEX_IMAGEGEN_SMOKE_TIMEOUT_SEC
• CODEX_IMAGEGEN_SMOKE_MAX_WAIT_MS
• CODEX_IMAGEGEN_SMOKE_POLL_INTERVAL_MS

Notes on text-heavy images

The service can submit any prompt, including infographic or timeline prompts. But if the model itself renders dense text poorly, the correct behavior for the skill is:

• warn the user
• keep the requested route
• let the user decide whether to proceed

It should not silently replace the request with an HTML or manual design workflow.

Publishing

See PUBLISHING.md for the release checklist and sanitization pass.

Naming note

Public project name:

• mindcode-claw-image2

Internal technical identifiers intentionally remain:

• codex-imagegen-service
• codex-imagegen-worker
• codex-imagegenctl
• CODEX_IMAGEGEN_*

⭐ Like it? Star it!

If you find this project interesting, give it a star, it helps others discover it too.

Follow Me

I write about AI, startups, and psychology.

• WeChat public account: MindCode
• X: @moneygalaxy
• Substack: mindcodeplus