Higgsfield, Krea, OpenArt and LUVI from Claude or ChatGPT: the MCP connectors compared
Higgsfield, Krea, OpenArt and LUVI Creator all run hosted MCP connectors that let Claude or ChatGPT generate on your account. They differ in whether you see the price before a run, how many models they expose, and whether the assistant gets each model's prompting guide.
An MCP connector lets an AI assistant such as Claude or ChatGPT generate images and video on your own account: you sign in once, then ask in plain language. As of September 25, 2026, four creator platforms run a hosted connector with account sign-in: Higgsfield, Krea, OpenArt and LUVI Creator. We read each one's own documentation on that date and compared what an assistant can actually do through it.
What does an MCP connector for AI generation do?
An MCP connector for AI generation is a hosted server that gives an assistant tools such as "list models", "generate" and "check status", acting on your account and spending your credits. MCP (Model Context Protocol) is the open standard Claude and ChatGPT use for these connections, so the assistant becomes the interface: you describe the shot, it picks a model, writes the prompt and returns the result.
What separates connectors is what happens around that generate call: whether you see the price first, whether the assistant knows how each model wants to be prompted, and whether it can chain several models.
Which platforms have a hosted connector?
All four sign you in with your existing account through OAuth, with no API key to create.
- Higgsfield. Endpoint
https://mcp.higgsfield.ai/mcp. Its help page lists Claude (claude.ai and Claude Desktop), ChatGPT and Cursor, plus agents through its CLI. Higgsfield's blog says the connector exposes "more than 30 video and image models". - Krea. Endpoint
https://api.krea.ai/mcp, with OAuth or an API token. Krea's product page names Claude, ChatGPT and Cursor; its developer docs list the tools, includinglist_models,get_model_schema,generateandget_upload_url. Neither page gives a total model count. - OpenArt. Endpoint
https://mcp.openart.ai/mcp. OpenArt names Claude, ChatGPT, Cursor and Kimi, and says "Generations through MCP spend the same credits as if you're generating on openart.ai." - LUVI Creator. Endpoint
https://www.luvicreator.com/api/mcp, for Claude and ChatGPT. 34 tools over 214 models for video, image, audio and 3D.
Do you see the price before the assistant generates?
Seeing the price before a run matters more through an assistant than in a web app, because the assistant may pick the model and settings for you. The four connectors handle it differently, as documented on September 25, 2026:
| Connector | Price before a run, as documented |
|---|---|
| Higgsfield | No automatic preview; ask the assistant for the cost |
| Krea | Not described in its MCP pages |
| OpenArt | Not described; charges "based on the model and resolution" |
| LUVI Creator | estimate_cost quotes the credits before a run |
Higgsfield's help page suggests a prompt for it: "Before generating anything, tell me the credit cost and wait for my confirmation." LUVI Creator's estimate_cost tool "works out how many credits a generation will cost, with the same pricing the charge uses", and the charge that follows can only go down from that quote. "Not described" means the documentation we read doesn't mention a quote, not that one cannot exist.
Does the assistant know how to prompt each model?
Each video and image model has its own prompt dialect, so an assistant that writes one generic prompt for every model loses quality on most of them. Krea's get_model_schema gives the assistant each model's inputs and their allowed values. LUVI Creator adds two tools on top of the schema: get_prompt_manual loads LUVI's prompting guide for the chosen model family, and check_prompt checks a drafted prompt against it and returns a corrected version before anything is spent. Higgsfield's and OpenArt's pages don't describe a per-model prompting guide.
Can the assistant chain several models?
Chaining several models, such as a product photo into a video ad, is where the connectors differ most. Krea exposes execute_node_app to run its node apps. LUVI Creator has eight workflow tools: the assistant builds and wires a workflow on the canvas, and estimate_workflow prices it. The run itself spends credits only if you allowed the assistant to run workflows with a credit ceiling and an expiry; otherwise it hands you a link to press Run yourself. Higgsfield's and OpenArt's pages describe single generations, account and project tools, not a multi-step chain.
Can you use your own images?
All four connectors accept your own files as references. Higgsfield and OpenArt open an upload window inside the conversation, Krea's get_upload_url returns a short-lived upload link, and LUVI Creator either gives a link that opens the upload dialog in your project or, when Claude can run code, uploads the attached file directly.
How we checked
- Sources: each platform's own MCP page, help article or developer docs, read on September 25, 2026 (linked below).
- LUVI Creator: tool names and descriptions from the public MCP documentation, model count from the models page, the same day.
- Results: No outputs were generated for this post, so it makes no claim about output quality.
- Limits: connectors change quickly. A feature missing from a vendor's documentation may still exist, and model counts are as each vendor states them.
Which one should I pick?
Pick by what you want the assistant to do. If you mainly generate on one platform you already pay for, its own connector is the shortest path. If you want the assistant to compare models across labs, see the price before each run and follow each model's prompting rules, those three things are documented in LUVI Creator's connector. To connect it, create a LUVI account, then follow the steps for Claude or ChatGPT.
Sources
- How to connect Higgsfield to Claude or ChatGPT — Higgsfield, help center, updated September 24, 2026. Accessed September 25, 2026.
- How to generate AI videos straight from Claude with Higgsfield's MCP — Higgsfield, blog, August 29, 2026. Accessed September 25, 2026.
- Model Context Protocol (MCP) server — Krea, developer docs. Accessed September 25, 2026.
- Krea MCP — Krea, product page, updated September 5, 2026. Accessed September 25, 2026.
- OpenArt MCP and CLI — OpenArt, product page. Accessed September 25, 2026.
Questions
What is an MCP connector for AI image and video generation?
A hosted server that an AI assistant such as Claude or ChatGPT connects to. After you sign in, the assistant can list models, generate and fetch results on your account, using your credits.
Which connectors show the price before generating?
Of the four checked on September 25, 2026, only LUVI Creator documents a tool that quotes the credit cost before a generation. Higgsfield's help page says to ask the assistant for the cost; Krea's and OpenArt's pages don't describe a quote.
Do I need an API key?
No. All four connectors sign you in with your own account through OAuth. Krea also accepts an API token for clients without OAuth.
More from Comparisons
- Six ways to make a face talk on LUVI, and the difference that decides the priceLUVI has six lip-sync families, and they do two different jobs: four animate a still portrait from an audio track, two re-drive the mouth in a video you already have. Which job you are doing also decides which clock you pay for—audio seconds or video seconds.
- The Negative Prompt has nearly vanished: how to exclude something without itFour of the 224 models in LUVI's catalog expose a Negative Prompt field, and they are all Qwen Image. The field disappeared because the sampling method it depends on did. Here is what replaces it on each family, and the three places where writing "no X" still works.
- Which LUVI video models generate sound—and the one line each of them needsMost video models on LUVI now render sound in the same pass as the picture, but they disagree about almost everything else: whether the switch is on, whether it costs more, and how you write the sound. Here is the audio switch and the audio syntax for each family, checked against their schemas.