MCP Server Overview

MeiGen helps you discover inspiration and create images and videos with the latest AI models in the AI tools you use. Install the ChatGPT plugin, connect through remote MCP, or use the open-source local server.

ChatGPT

Open MeiGen in the official ChatGPT store and select Install.
Sign in to your MeiGen account and complete the connection. No API key or Node.js is needed.
Select or mention MeiGen in a conversation, then ask it to find inspiration or create an image or video.

Try: “Use MeiGen to find minimalist product photography inspiration. Show me reference images and their prompts.” Search does not spend generation credits. Generation uses your purchased MeiGen credits; daily free credits do not apply. Review the model, image count and total price before generating. The ChatGPT plugin and remote OAuth connection use MeiGen Cloud. Local npm remains available for existing MeiGen API-key integrations and explicitly selected external-provider BYOK. Version 3.0.0 removes ComfyUI and the bundled offline prompt library; see the migration notes.

GitHub Repository

View source code, report issues, and contribute.

Compose inside another workflow

Keep the upstream script, prompt, model/provider, count and approved budget. Creative helpers are optional. Use persistent UUID requestId, wait: false, structured results and check_generation to submit frames/videos and resume them without re-approving every step. The caller owns bounded scheduling, previews and downloads. See the complete workflow contract and examples.

What is MCP?

Model Context Protocol (MCP) is an open standard that allows AI assistants to use external tools. When you install MeiGen as an MCP server, your AI assistant gains the ability to generate images and videos, run semantic search across the MeiGen community prompt gallery, and enhance prompts — all through natural language. Standard remote MCP and local npm support anonymous public search; the official ChatGPT plugin requires account authorization for search. Prompt enhancement is local. In 3.0.0, inspiration search uses the online gallery only, without an offline fallback; unavailable services return an error.

Five dedicated image workflows

Ask for a cutout, product detail images, a marketing poster, a new background or image enhancement. The agent should inspect list_skills, prepare the actual materials, submit the matching tool and use check_skill for results. It generates request UUIDs itself. A specified count/quality is already your instruction; only unresolved scope or an Upscale warning needs a decision.

Workflow	Required material	Optional material	Output
Remove Background	One source image	None	One transparent PNG
Product Detail Images	One product photo and selected modules/count	Product name, verified selling points, logo, model image, two extra angles	1–6 paid images
Marketing Poster	Brand, event or topic	Copy, logo, three product images, style reference	One paid poster
AI Backgrounds	One product photo	Mode; custom mode requires a background description	One paid image
Upscale	One original still JPEG/PNG/WebP	Crisp/creative mode	One paid enhanced image

Skills use purchased credits only, with no free attempts or daily credits. BYOK do not supply these workflows. See the complete Skills API and material guide. For posters, put visible copy in content; use extraNotes or customStyle for design directions. autoCopy: false preserves supplied wording, with selected-language translation still possible. See poster field roles and a complete MCP call.

Choose a connection

ChatGPT users should install the official plugin. Other new users should connect with remote OAuth at https://www.meigen.ai/api/mcp, then sign in to MeiGen and approve access in the host. No API key or Node.js is needed. Use a compatible host if OAuth is unavailable. Local npm retains existing API-key compatibility for generation, public search and Skills, BYOK, local file preparation, prompt enhancement and local preferences. It has no OAuth login or private-library tools. API keys are not offered as another new-user authentication method. OpenClaw users can install the ClawHub plugin and connect with OAuth. Existing local Claude/OpenClaw connections remain a separate compatibility path; avoid duplicate tools when switching. The 3.0.0 local image providers are MeiGen (existing legacy key) and OpenAI-compatible APIs, in that priority order. ComfyUI is removed. Skills and video remain MeiGen Cloud features; use list_models and list_skills for current models and prices.

Tools

These 16 shared tools include reference-media preparation. Local npm 3.3.0 has 18 tools, including two local helpers. Standard remote MCP and the official ChatGPT adapter each expose 23 tools, adding credit balance, saved characters and the five private-library tools. Refresh cached tool discovery after an update; access still depends on the host and the permissions you approve.

Tool	Purpose	MeiGen account authorization
`generate_image`	General image generation; local can use configured backends	For MeiGen Cloud
`generate_video`	MeiGen Cloud video generation, including opt-in drafts	Yes
`render_video_draft`	Separately confirmed 1080p final from a completed draft	Yes
`check_generation`	Status/recovery by generationId or requestId	Request-ID lookup requires the owning account
`list_models`	Available models and capabilities	No
`search_gallery`	Search public prompts and image previews; batch limits vary below	Standard MCP/local: optional; official ChatGPT plugin: required
`get_inspiration`	Read a public gallery entry	No
`list_skills`	Five workflows, materials, live prices and input schemas	No
`upload_skill_image`	Prepare real image URLs or attachment bytes	Yes
`upload_reference_media`	Prepare a selected video/audio URL for a specific model	Yes
`remove_background`	One transparent cutout	Yes
`generate_product_detail_images`	1–6 paid product modules	Yes
`generate_marketing_poster`	One promotional poster	Yes
`generate_ai_background`	One white, smart or custom background	Yes
`upscale_image`	Enhance one original still image	Yes
`check_skill`	Private Skill receipt, module status and recovery	Yes, owning account; legacy local calls require the original key

Local additionally provides enhance_prompt and manage_preferences without a MeiGen account. Version 3.0.0 removes comfyui_workflow and offline prompt fallback.

Public gallery search	Default / maximum results per batch	Account access
Standard remote MCP	3 / 3	Optional OAuth; anonymous searches are IP-limited
Local npm 3.3.0	3 / 3	Anonymous, or an existing legacy key
Official ChatGPT plugin	6 / 6 after tool refresh	OAuth with search permission (`meigen:search`)

Authorized searches use the account’s daily search allowance, not generation credits. Starting a search uses that allowance; continuation pages do not count again. Standard remote MCP and local 3.0.0 support pagination; the official ChatGPT adapter also offers a next-batch card when the host loads its current tools. To see another batch, preserve the returned query and snapshot ID and pass nextOffset as offset; stop when hasMore is false. A different direction starts a new query. Keep original generation request IDs during recovery; a timeout is not permission for a new paid task.

Midjourney controls

generate_image supports these controls on standard remote MCP, the official ChatGPT plugin, and local npm 3.3.0 or later with the MeiGen provider. Refresh hosted tools or update the local package pin and restart. Use model ID midjourney-v8.1 for Midjourney V8.2; the stable ID has not changed.

Input	Use
`referenceImages`	Prepare the selected reference with `upload_skill_image` and pass its returned URL. Local npm can also upload a host-readable file. Follow the live model reference limit.
`referenceType`	`content` (default) or `style`. A supplied style image takes precedence over `niji7Options.sref`; saved characters remain content references.
`niji7Options`	`stylize` (0–1000), `chaos` (0–100), `raw` (boolean), `iw` (0–3), `sref`, `sw` (0–1000), `profile`, and MJ `quality` (`"1"` or `"4"`). V8.2 always uses `sv=6` and ignores `weird`.
`resolution`	`1K` or `2K`. Each request returns four candidates, billed once: currently 12 credits at 1K or 18 at 2K. Quote the live `list_models` price before submission.

For example, these are ordinary generate_image tool arguments; submitting them is a paid action, not a connection check. Replace the sample requestId with a new persisted UUID for each newly authorized job:

{
  "requestId": "03c08bc1-d158-40af-a667-7ba901788a7f",
  "modelId": "midjourney-v8.1",
  "prompt": "A ceramic coffee mug in soft window light",
  "aspectRatio": "2:3",
  "resolution": "2K",
  "niji7Options": { "stylize": 100, "raw": false, "quality": "4" }
}

MJ quality belongs inside niji7Options; top-level quality is a separate image-model setting. Prefer structured options over prompt flags. sref accepts numeric codes, random, or image URLs; profile accepts alphanumeric codes separated by spaces or commas. The official ChatGPT plugin requires direct images.meigen.ai asset URLs for reference images and image-valued sref; upload external images first. Standard remote MCP and local npm accept public HTTPS sref image URLs. Local file paths belong in local npm referenceImages, never sref. Remote saved characters require the existing character-read permission and exact referenceToken from list_characters; check capabilities.characters and reference limits before combining inputs. Prompt translation and long-prompt compression continue in the backend. During recovery, keep the original requestId, reference role, and all option values unchanged. See the full parameter reference.

Private library through remote OAuth

Standard remote MCP provides search_favorites, list_boards, search_history, get_library_item and save_favorite. Connect with OAuth and approve the separate library read/write permissions when requested; existing connections do not gain them automatically. Local npm and API keys cannot access these tools. The official ChatGPT adapter provides the same library functions; reconnect or refresh its tools if they are missing. Lists contain prompt excerpts of at most 500 Unicode code points. Use get_library_item for a selected user prompt in pages of 8000 code points by default, up to 16000; internal Skill prompts are never returned. Saving requires an explicit request or prior opt-in, preserves existing notes on repeat calls and only supports owned private board destinations. Favorites are context for preferences, not a replacement for current instructions.

Video capabilities differ by model — quality tiers, available durations and resolutions, audio, reference-video continuation, and whether a first frame is required. Some models are image-to-video only. Call list_models for the current lineup, or see the model comparison.Seedance 2.5 uses model ID seedance-2-5, supports 4–30 second 480p/720p/1080p output and reference-video continuation, and has no tier option.generate_video requires an explicit model. Use firstFrame / lastFrame for endpoint frames and contentReferenceImages for content, identity or style. Pass contentReferenceImages: [] for endpoint-only requests. Follow list_models.capabilities.videoInputs and its tier overrides; do not mix first/last frames with content references or saved characters when the model forbids it.Use upload_skill_image for images and upload_reference_media for selected direct HTTPS video/audio URLs, then pass the returned MeiGen URL to referenceVideos / referenceAudios. The official plugin also supports accessible host attachments. Remote tools cannot open local paths; existing local npm integrations can still prepare local files. Preparation needs a positive purchased balance but does not deduct credits. Prepared video/audio links are public and temporary, with cleanup scheduled seven days after upload; reuse within that period and do not use this as permanent hosting.Use the live pricingMode and capabilities.video.billing to quote a job. Some models bill input-video plus output seconds; output_only bills only output. Audio seconds are not added, but audio or content images can select a different reference rate on models that advertise it. HeyGen is available; reference-video availability follows the live catalog. Honor capabilities.video.referenceAudio.canCombineWithVideo: when false, do not combine separate audio with video. Embedded audio is allowed. HeyGen bills output seconds plus each reference video’s rounded-up seconds at the reference rate.

Balance and saved characters

Remote OAuth provides get_credit_balance for purchased credits available to MCP; daily free credits are excluded. list_characters reads existing saved characters with separate meigen:characters:read consent. Use the exact returned referenceToken in a supported model’s prompt, check capabilities.characters, and use the model’s referenceMentionSyntax for uploaded media. Character creation is not supported. Local npm and API keys cannot read or use private saved characters.

GPT Image 2.5

Call list_models first to check availability. Local npm 3.0.0 supports model: "gpt-image-2.5" or its modelId alias, quality (low, medium, high, xhigh, max) and modelVariant: "sunburst" (default) or "flare". If both model and modelId are supplied, they must match. The remote MCP server and REST API use modelId with modelVariant. GPT Image 2 remains a separate model with its original pricing.

Next steps

Remote OAuth setup for Claude, Cursor and compatible clients.
Legacy local setup and BYOK for existing integrations.
3.0.0 migration notes for the ComfyUI/offline-library removal.

Video drafts

Seedance 2.5 can offer a paid 480p draft followed by a separately authorized 1080p final. Availability depends on the connected tools; see drafts, pricing and recovery.

Getting Started

Features

MCP Plugin

Account

ChatGPT

GitHub Repository

Compose inside another workflow

What is MCP?

Five dedicated image workflows

Choose a connection

Tools

Midjourney controls

Private library through remote OAuth

Balance and saved characters

GPT Image 2.5

Next steps

Video drafts

​ChatGPT

GitHub Repository

​Compose inside another workflow

​What is MCP?

​Five dedicated image workflows

​Choose a connection

​Tools

​Midjourney controls

​Private library through remote OAuth

​Balance and saved characters

​GPT Image 2.5

​Next steps

​Video drafts

ChatGPT

Compose inside another workflow

What is MCP?

Five dedicated image workflows

Choose a connection

Tools

Midjourney controls

Private library through remote OAuth

Balance and saved characters

GPT Image 2.5

Next steps

Video drafts