Higgsfield MCP is a hosted Model Context Protocol server at https://mcp.higgsfield.ai/mcp that gives an AI agent direct access to Higgsfield's image and video models. You add it as a connector, sign in with your Higgsfield account, and your agent can generate images, create videos, train characters and browse your generation history inside a single session. There is no API key to manage.
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, in three steps
Higgsfield documents the Claude flow directly:
- Copy the connector URL —
https://mcp.higgsfield.ai/mcp - Go to Claude → Customize → Connectors. In Claude desktop or claude.ai, add a custom connector, name it
Higgsfield, and paste the URL. - Connect and sign in. Authenticate through your Higgsfield account, then ask the agent to generate an image or video.
If you are on Claude Code or Codex, Higgsfield recommends its CLI instead of the MCP connector — it is a separate tool, open-sourced at github.com/higgsfield-ai/cli.
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.
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.