You connect Higgsfield to Claude with one custom connector. Open Claude, point it at Higgsfield's MCP endpoint, sign in once, and Claude can generate images and video straight from the chat. No API keys, no glue code. The whole thing takes about two minutes.
Higgsfield shipped its MCP connector on 30 April 2026. It exposes 30+ models (Soul, Nano Banana, Seedance, Kling, Veo, and more) to Claude on web, desktop, and Claude Code. Once it's wired up, you ask Claude for an image or a clip in plain language and it picks the model, runs the job, and drops the result back in the thread.
This guide is the fast path: connect, authenticate, run a first test. If you want the full breakdown of every model family and how to keep credits from evaporating, that lives in the full setup guide (linked at the end). Here we just get you live.
Before you start
You need a Higgsfield account with credits. There's no free plan as of June 2026: Plus runs around $40/mo, with Business and Ultra above that. Confirm the current numbers on higgsfield.ai/pricing before you commit, since the tiers move. I ran everything below on a Plus plan.
Pick your path. Most people are on Claude web or desktop, so start there. If you live in the terminal, the Claude Code path is below it and it's even shorter.
Path 1: Claude web or desktop
This is a one-time setup. After it's done, Higgsfield shows up in every new chat.
1. Open Claude and go to Settings, then Connectors.
2. Click Add custom connector.
3. Name it Higgsfield.
4. Paste the MCP URL:
https://higgsfield.ai/mcp5. Save, then sign in to your Higgsfield account when the browser prompt opens.
6. Approve the connection. Higgsfield's tools now appear in Claude.
Last step worth doing now: open the connector's tool permissions and decide what Claude can run without asking. Leave it on confirm-first and Claude pauses for approval before each generation. Flip the tools you trust to auto-allow and it just runs them. For a first test I'd leave it on confirm-first so you see exactly what fires. The full treatment of that setting is in the setup guide.
Path 2: Claude Code
One command. It uses HTTP transport and browser OAuth, so there's no API key to paste anywhere.
claude mcp add --transport http --scope user higgsfield https://mcp.higgsfield.ai/mcpRun it, and Claude Code opens a browser tab to authorize. Sign in to Higgsfield, approve, and you're connected at the user scope (available across your projects). That's the entire setup.
Note the endpoints differ slightly: the web connector uses higgsfield.ai/mcp, Claude Code uses mcp.higgsfield.ai/mcp. Use the one that matches your path.
Run your first generation
Connected. Now prove it works. Start with an image, because every Higgsfield video model needs a start frame anyway. There is no text-only video path: you generate a still first, then animate it.
Type something like this into the chat:
Generate a 16:9 product photo of wireless earbuds in their case on a marble surface, soft daylight.Claude picks a model (it reached for Nano Banana on mine), runs the job, and posts the image back in 10 to 20 seconds. If you left permissions on confirm-first, you'll approve the tool call before it fires.
Once you have an image you like, animate it. Ask Claude to turn that still into a short clip and it routes to a video model and uses your image as the start frame. Here's the first clip I generated right after connecting, a talking-creator vertical with generated audio:
That's the loop: ask, approve, get an asset. From a cold connector to a finished video clip in one sitting.
What to watch for
A few things will trip you up early, so know them going in.
Video is image-to-video only. You can't prompt a clip from text alone. Generate or upload an image first, then animate. Claude handles the routing, but the start frame is on you.
External image URLs are flaky. Pointing a job at an image hosted elsewhere can throw a hosting error. Generating the image inside the flow, or uploading it directly, is far more reliable.
Generated audio sometimes drops. A clip occasionally comes back silent when it should have audio. Re-run it and it usually lands the second time.
Credits go fast. This is the real cost. On Plus I started at 787.75 credits and was at 734.98 after five images and two videos, so roughly 53 credits for seven assets. Call it 7 to 8 credits per still, a bit more per short clip. Generation is quick enough that an unplanned afternoon can burn a chunk of your month. The fix is a draft-first habit (short, low-res passes before the full render) and asking Claude to check your balance mid-chat. The full credit-budget system is in the setup guide.
Where to go next
That's the connection done. If you want the complete picture, every model family, the auto-confirm setting in depth, and the system I use to keep credit burn under control, read the full Higgsfield MCP setup guide at /blog/higgsfield-mcp-setup-guide. And if you're still deciding whether Higgsfield earns the subscription, the hands-on review with real generated proof is at /tools/higgsfield.
Vibetoolstack reviews tools we'd recommend to readers building toward $10k/mo of independent income. Where an affiliate program exists and we participate, the link is marked. Where not, links are editorial. The verdict above doesn't depend on affiliate status.