Higgsfield MCP is a hosted Model Context Protocol server at https://mcp.higgsfield.ai/mcp. As of October 2026 you add it to Claude (Settings → Connectors), ChatGPT (Plugins Directory) or Cursor (Marketplace), sign in with your Higgsfield account, and generate images and video from chat. It requires an active subscription, and every generation deducts credits.
What changed (3 October 2026): The Claude menu path is now Settings → Connectors (it was Customize → Connectors), ChatGPT and Cursor routes are documented, and MCP requires an active Higgsfield subscription. Higgsfield also launched a separate pay-per-generation API on 15 September 2026; it does not use MCP or plan credits. Source: Higgsfield's connect-an-agent help page, read 2026-10-03.
The part almost nobody mentions: generations made through MCP always deduct credits, even on plans that advertise Unlimited access. More on that below, because it is the difference between a workflow that costs nothing and one that quietly drains your balance.
Setup
You need an active Higgsfield subscription and an account in your agent. There is no API key. Per Higgsfield's help centre:
Claude (claude.ai or Claude Desktop)
- Go to Settings → Connectors and choose Add custom connector.
- Name it
Higgsfieldand pastehttps://mcp.higgsfield.ai/mcp. - Click Connect, sign in to your Higgsfield account and authorise.
ChatGPT: open the Plugins Directory (or higgsfield.ai/mcp), add the Higgsfield plugin and sign in. Audio generation and the Website Building skill are not available in ChatGPT.
Cursor: Customize → Marketplace → Higgsfield → Add, then sign in.
CLI agents (Claude Code, OpenClaw, Hermes): run npm i -g @higgsfield/cli, then higgsfield auth login, then npx skills add higgsfield-ai/skills. For Claude Code or Codex Higgsfield recommends the CLI over the MCP connector; it is open source at github.com/higgsfield-ai/cli.
If auth expires, disconnect and reconnect. If the agent does not call Higgsfield, start a new conversation and name Higgsfield explicitly.
Which agents actually work
Higgsfield lists official support for Claude (web, Cowork, and Claude Code), OpenClaw, Hermes Agent, and NemoClaw. Because MCP is an open standard, any MCP-compatible client can connect — including a self-hosted agent on your own VPS or local machine. You point it at the same server URL.
What models you get
Higgsfield says the connector exposes 30+ models. The ones it names explicitly:
| Model | Type |
|---|---|
| Nano Banana Pro | Image |
| GPT Image 2 | Image |
| Seedream 5.0 Lite | Image |
| Soul 2.0 | Image (character-consistent) |
| Seedance 2.5 / Seedance 2.0 | Video |
| Kling 3 | Video |
| Google Omni Flash | Multimodal |
| Veo, Flux, Minimax Hailuo, Cinema Studio | Video / image |
Output limits: images up to 4K resolution, videos up to 15 seconds. Your agent picks a model automatically based on the task, or you can name one in the prompt.
Generation runs asynchronously — the agent submits the job and polls for the result rather than blocking, which is why a long video request does not time out your session.
The credit rule that catches people
This is the single most useful thing to know before you wire MCP into a workflow.
Higgsfield's credits documentation is explicit: credits are deducted for all generations on web, MCP, CLI, Canvas, and Supercomputer — including images, videos, re-rolls and upscaling. What does not deduct credits is Unlimited model generations and the free generations included with your plan.
But — and this is the catch — Unlimited access and free generations apply only on higgsfield.ai. Any generation made outside it, including through MCP, CLI, Canvas and Supercomputer, always deducts credits, regardless of plan.
So if you bought a plan for its Unlimited models and then moved your workflow into an agent, you are now paying credits per generation for something that was free in the browser. Two practical consequences:
- Prototype in the browser, automate in MCP. Burn the free/Unlimited generations on iteration, then use the agent for the runs that matter.
- Watch the balance during agent loops. An agent that re-rolls a video five times to satisfy a prompt has spent five generations' worth of credits, and there is no per-generation confirmation step the way there is in the UI.
Credits from genuinely failed generations are returned automatically — you can confirm that under Manage Account → Usage, which logs every operation with the model used and credits deducted.
What it is actually good for
The honest answer is agent-driven pipelines where the generation is one step of many. Higgsfield's own examples are a full UGC flow — concept, script, creator performance, product close-up, captions and a social-ready cut, produced in one session — plus faceless video, motion graphics, and website builds from a reference image.
That framing matters. MCP is not a faster way to make one video; the browser is fine for that. It is a way to put generation inside a loop that also writes the script, checks the brief and assembles the output.
MCP, CLI or API?
MCP and CLI attach an agent to your existing plan and spend its credits (plan prices; what is free). If you want to call Higgsfield from your own product, the Higgsfield API bills a separate US dollar balance instead.
Where this fits if you are making ads
If your end goal is UGC ad creative rather than general video, the generation model is rarely the bottleneck — the script and the hook are. Creetr's UGC script generator, hook generator and creative brief generator run free with no signup and no credit meter, so you can settle the concept before you spend a single Higgsfield credit on rendering it.