MeiGen AI Design — Claude Code Plugin

AI image and video generation with guided ecommerce Skills, creative workflows, parallel general image generation, and a 1,446-entry inspiration library. See the full guides in English or 中文.

Using Codex or ChatGPT? See the Codex setup, ChatGPT web support, or copy the AI installation prompt in the main guide.

Callable tools and optional creative help

The caller owns scripts, prompts, models/providers, ratios, approved count/budget, scheduling and presentation. Invoke tools directly from another Skill or agent; creative planning and the plugin agents are optional. They preserve resolved input and do not reconfirm each authorized frame, clip or batch. Discovery can support any task, and actual visual inspection is allowed before making visual claims.

For composed MeiGen jobs, persist UUID requestId and exact inputs, use wait: false, download: false on local npm, and resume via check_generation(requestId=...) or the returned generationId. Existing synchronous/download defaults remain enabled; remote MCP returns URLs without a download parameter. The caller reads structured results and manages in-flight budget reservations. Local npm bounds API submissions at four and runs ComfyUI serially; actual backend limits remain authoritative. No blanket video-serialization or ten-image rule applies. See workflow examples.

Connect an MCP Server

Choose one connection for this host. Remote MCP provides 14 tools for MeiGen generation, gallery and Skills. Local npm MCP 2.0.2 provides 17, adding prompt enhancement, local preferences and ComfyUI management. The local connection is needed for those plugin features and automatic local file preparation; remote MCP is the simplest option for cloud Skills.

Remote — no npm install

In the host's MCP settings, choose Streamable HTTP, set the URL to https://www.meigen.ai/api/mcp, and configure the header Authorization: Bearer YOUR_MEIGEN_API_KEY. Claude Code also supports:

# MEIGEN_API_TOKEN must already be set in this terminal's environment.
claude mcp add --transport http meigen https://www.meigen.ai/api/mcp \
  --header "Authorization: Bearer $MEIGEN_API_TOKEN"

The server uses stateless Streamable HTTP. Accepted generation jobs and billing records remain on the server for recovery. Remote tool updates need a backend deployment, not an npm release; refresh or reconnect if the host caches tools. Remote generation requires internet access. Plugin instructions and local npm behavior still require their own updates when changed.

Local — all plugin features

The official MeiGen marketplace plugin already bundles this connection. The manual example below is for a bare MCP install or a distribution without MCP wiring, such as the separate wshobson marketplace. Keep one entry.

Node.js 22 or newer is recommended. Add this server to your project's .mcp.json; store the key in local connection settings rather than committing it:

{
  "mcpServers": {
    "meigen": {
      "command": "npx",
      "args": ["-y", "meigen@2.0.2"]
    }
  }
}

Pass MEIGEN_API_TOKEN in this server's environment for MeiGen generation and Skills. Restart or reconnect Claude Code after installation or configuration changes. Do not configure both entries under different names unless you intentionally want duplicate tools.

First-Time MeiGen Setup

1. Sign in and open API Keys in a desktop browser. Create a key beginning with meigen_sk_; the mobile site currently redirects this page.
2. On the same account, open Profile → Top Up or mobile Premium to purchase credits.
3. Enter the key in the MCP connection's credentials settings: the Authorization header for remote MCP, or MEIGEN_API_TOKEN for local npm. Never paste a key into chat.
4. Ask “Show available Skills and their current prices.” list_skills works without a key and does not generate an image or spend generation credits.

API generation uses purchased credits only. Daily free credits and Web free attempts do not apply. The five Skills require MeiGen Cloud even if OpenAI or ComfyUI is configured.

Choose a Skill

For a direct user request, the assistant chooses the Skill, prepares images and follows status; an upstream caller receives structured results and controls presentation. It should use the user's specified count or modules without asking for the same confirmation again. If product-detail count is unresolved, clarify it. MCP requires an explicit modules selection; only direct HTTP calls default to three paid images (hero/detail/scene) when modules are omitted. Use live list_skills prices to calculate the batch cost; do not add unrequested images. Ask to preserve exact wording when supplying final copy.

For Upscale, pass the original file/URL directly or use upload_skill_image with purpose: "upscale". Originals may be up to 64 MiB / 64 million pixels; local encoding must fit 9,500,000 bytes without resizing (otherwise use a public original URL). Base64 remains limited to 3 MiB decoded. For upscale_resize_required or price_changed, explain the change and obtain acceptance before using a new requestId and the accepted flags/price. See the full limits.

Local npm accepts real absolute local file paths and public direct HTTPS image URLs. For remote MCP, use upload_skill_image with a public direct URL (up to 8 MiB) or actual readable attachment bytes (up to 3 MiB decoded), then pass the returned URL to the Skill. Existing supported MeiGen image URLs can be used directly. Private-network sources, redirects, login-protected links and IPv6-only sources are unsupported. Uploading prepares a reference and does not spend generation credits. If the host cannot read an attachment, use a direct image link or local npm.

New paid requests get a client-generated requestId. After interruption, use check_skill with the original key, Skill and ID; follow nextAction and preserve returned retryParameters exactly. Typical polling is every 10 seconds. Display completed modules and their resource links; report failures/refund states individually. Do not ask users for UUIDs or automatically submit a new paid batch to replace failed modules.

Other Providers for Local General Image Generation

generate_image in the local npm server also supports the following providers. They do not replace MeiGen Cloud for the five Skills or for MeiGen video generation. The plugin's /meigen:setup command can explain provider settings; keep credentials in local connection settings/configuration.

Local ComfyUI

1. Install and start ComfyUI.
2. Set COMFYUI_URL to the server URL, for example http://localhost:8188.
3. Export a ComfyUI workflow in API JSON format and import it through the plugin.

With local files and a local workflow, generation can stay on your machine. Gallery searches and MeiGen Skills still use external services.

OpenAI-compatible API

Configure OPENAI_API_KEY, OPENAI_BASE_URL and OPENAI_MODEL for your image provider in the MCP server's environment. Use a model and endpoint supported by that provider; it handles its own billing. These settings apply to local general image generation, not Skills.

Agents

These optional plugin agents support general image workflows; direct tool calls are equally supported. The five Skills perform their own planning and do not require them.

Auto-preview is off by default. To open saved images on macOS, explicitly set MEIGEN_AUTO_OPEN=1 in the plugin host environment. Workflow callers otherwise own presentation.

Commands

These commands belong to the Claude Code plugin; a bare MCP connection does not install them.

Troubleshooting

OpenClaw distribution

The OpenClaw manifest explicitly declares its MCP server and skills. Some CLI versions load this directory as a Claude-compatible bundle instead; Claude-only commands, agents, hooks and output styles depend on the actual loader and are not promised by the native manifest. The standalone ClawHub Skill is a separate distribution. See the upgrade guide and direct HTTP API.

Publishing

For maintainers, see RELEASING.md. NPM_TOKEN belongs in the repository's ignored .env.local for publishing; it is separate from the customer's MEIGEN_API_TOKEN and must not be committed or packaged. npm 2.0.2 is the package version, independent of the MCP protocol date and SDK version.

License

MIT