If you have seen Higgsfield MCP mentioned in the last few weeks, you probably saw it framed as a shortcut: connect it to Claude or ChatGPT, type a sentence, get a finished video. That part is true. What almost nobody mentions is the line Higgsfield prints on its own pricing page, directly under the plan cards.
Unlimited models and Free Generations on plans are accessible only via higgsfield.ai and are not accessible on MCP/CLI, Canvas or Supercomputer.
Read that again if you are on an unlimited plan. The thing you paid for does not follow you into the MCP. This is not a bug or a rollout gap - it is documented policy, and it is common enough that Higgsfield answers it in its own FAQ. This post covers what the MCP genuinely does well, what it costs in practice, and who should skip it.
Higgsfield AI's MCP is a hosted Model Context Protocol server run by Higgsfield itself, at https://mcp.higgsfield.ai/mcp. It is not the Higgsfield API, which is a separate developer product. Higgsfield's help center puts it plainly: "It isn't the Higgsfield API: MCP connects to your existing account and uses your existing plan credits, with no API key required." API keys belong to that developer API, which Higgsfield points to for building generation into your own product with per-request billing.
That last detail matters more than it sounds. You authenticate with OAuth, the same way you would log into any site, so no API key ever gets pasted into a config file. As of October 12, 2026, an unauthenticated request to the endpoint gets a 401 with an OAuth metadata header pointing at a standard authorization flow - which is the boring, correct behavior you want from something holding your billing account.
One thing to watch: third-party wrappers with similar names exist on GitHub and in MCP directories, and at least one of them predates Higgsfield's own CLI repo by about six months. Higgsfield's own server is not listed in the official MCP registry, so you cannot use registry presence as a trust signal here. Use the mcp.higgsfield.ai URL from Higgsfield's own documentation and nothing else.
Which clients it connects to
As of October 12, 2026, Higgsfield's MCP landing page has setup tabs for ChatGPT Plugin, Claude, Claude Code, CLI and Other, and its page description names Claude, OpenClaw, Hermes Agent and NemoClaw alongside "any MCP-compatible client". Cursor is no longer on the landing page, but the help center still documents a Cursor setup. The connection method differs by client, and the difference is not cosmetic.
Claude (claude.ai and Desktop): Settings, Connectors, Add custom connector, paste the MCP URL, sign in.
ChatGPT: there is an official Higgsfield plugin in the Plugins Directory. You add the plugin rather than typing an MCP URL.
Cursor: open Customize, then Marketplace, find Higgsfield and click Add.
Claude Code, OpenClaw, Hermes: these agents connect through the Higgsfield CLI, which the agent installs itself with npm i -g @higgsfield/cli, then higgsfield auth login.
Any other MCP-compatible client: paste the same server URL into the agent's settings.
If your Claude account is managed by an organization, an admin has to allowlist the Higgsfield MCP URL before the connector will appear. That trips up a lot of people on work accounts.
The credit rule that changes the math
Here is the part that deserves a full section rather than a footnote. Higgsfield's help center publishes a direct comparison between the web app and the MCP:
Higgsfield's own help center table. Unlimited access and free generations are web-only (captured 13 September 2026)
The row for unlimited model access reads "Yes, where included in plan" for the web and "No, always deducts credits" for MCP. Free generations: the same split. The FAQ entry is almost comic in how direct it is - "I have Unlimited access to a model, but MCP charged credits for it. Is that a bug?" answered with "No, that's expected."
The same rule applies to the CLI, Canvas, Supercomputer and, in the help center's words, "other automated tools". Anything automated bills at standard credit rates.
The same policy appears on the pricing page itself, under the plan cards (captured 13 September 2026)
What that costs in real numbers
Credit prices come from Higgsfield's own pricing page and its Seedance pricing post. A Seedance 2.5 clip at 720p and 8 seconds costs 52 credits. Seedance 2.0 at the same settings costs 36. The plan-picker on the pricing page prices a Kling 3.0 generation (8 seconds, 720p) at roughly 14 credits and a Nano Banana Pro image at 2.
Set that against what each plan actually includes per month, as listed on the pricing page as of October 12, 2026:
Higgsfield's own plan picker prices a Kling 3.0 clip at about 14 credits and a Nano Banana Pro image at 2. The Plus card beside it promises 5,000 free Soul 2.0 generations, none of which apply through MCP (captured 13 September 2026)
Plan
Price
Credits/mo
Seedance 2.5 clips via MCP (720p, 8s)
Kling 3.0 clips via MCP (720p, 8s)
Starter
$19/mo
270
not on this plan
about 19
Plus
$59/mo, or $47/mo billed annually
1,200
about 23
about 85
Ultra
$129/mo, or $99/mo billed annually
3,000
about 57
about 214
The plan cards show higher estimates, about 80 Seedance 2.5 videos on Plus and about 200 on Ultra. Those fit shorter 480p clips rather than 720p, 8-second ones: Higgsfield's Seedance pricing post puts an 8-second clip at 24 credits at 480p.
So the honest summary of a Plus plan is this: on higgsfield.ai the plan card promises seven unlimited models and 5,000 free Soul 2.0 generations. Through the MCP none of that applies - you get about 23 Seedance 2.5 clips at 720p and 8 seconds, and then you are buying credit packs. Both of those sentences describe the same $59 subscription.
Starter is the one to think hardest about. It has no Seedance 2.5 access at all, so an agent on Starter is working with the cheaper models, and 270 credits does not stretch far even there - on the models Starter can reach that is roughly 19 Kling 3.0 clips, and that is the whole month. If your plan for the MCP is "let Claude generate videos for me on demand", Starter will not carry it.
The feature nobody uses but should
The MCP's generation tools accept a get_cost flag. Set it to true and the server returns the credit cost of the job without submitting it. In an agent workflow where a single vague instruction can spawn four variants, each billed separately, a dry run before the real call is the difference between a controlled spend and a surprise.
Worth knowing alongside it: if a request times out, do not let an agent retry automatically. The submission may have gone through, and a blind retry bills you twice. Failed jobs usually get refunded, but duplicates are not failures.
How it compares to other video MCPs
Higgsfield is not the only video platform with an MCP server, but the field is thinner than you would expect.
Runway has a hosted server at mcp.runwayml.com plus a self-host repo, and like Higgsfield it needs no separate API key for the hosted version.
Replicate has a hosted OAuth server and far more models, but they are raw models rather than finished workflows.
fal.ai also connects over OAuth: its docs cover ChatGPT (through the fal plugin), Claude, Claude Code, Codex CLI and Cursor, all pointed at https://mcp.fal.ai/mcp-relay with browser sign-in and no API key. Its catalog spans image, video, audio, speech, 3D and language models, but like Replicate it gives you models rather than finished workflows.
Kling has an official server and CLI, added via claude mcp add.
Luma has a repo in its GitHub org with a single commit and no announcement. Treat it as experimental.
Sora and Veo have no first-party media-generation MCP. OpenAI's official MCP is documentation-only and states it does not call the OpenAI API on your behalf.
Higgsfield's real differentiator is not catalog breadth: its MCP page advertises "30+ models for image and video generation", while Replicate and fal run general-purpose catalogs. It is that the MCP layers workflows on top of generation - Soul characters, preset skills for marketing, UGC and faceless videos, and, as of October 12, 2026, skill bundles on the MCP page such as Paid Ads, Motion Design and a Production bundle with 11 workflows for 3D, VFX, editing and design. If you want an agent to run a pipeline rather than call one model, that bundle is the argument.
Pricing and what you need before connecting
The three individual plans on 13 September 2026: 270, 1,200 and 3,000 credits a month. Annual billing saves nothing on Starter, and Starter is marked as having no Seedance 2.5 access
There is no trial to lean on. As of October 12, 2026, Higgsfield's connection guide says "An active paid subscription is required", and its plans article answers the free-trial question with "There's no trial period." The free tier covers a small set of models but "doesn't include credits", and every MCP generation deducts credits, so the free tier is not a way in. The refund policy is not a trial either: refunds apply within 7 days of the initial purchase only if no credits have been used, with a service fee of up to 6%.
The practical way to test the MCP is the smallest plan that covers the models you need. Seedance 2.5 needs Plus or above, since the Starter card lists only Seedance 2.0 Fast and 2.0 Mini. Before the first real job, tell the agent to quote the credit cost and wait for your confirmation, which is the instruction Higgsfield's own guide suggests at the start of a session.
If you do subscribe, Higgsfield's affiliate page says people who log in through a partner link get a personal discount of up to 50%, active for three hours from login. The discount is personal, so it may be lower than 50%. The link below is our affiliate link.
Does the MCP cost extra on top of my subscription?
No. There is no MCP fee and no API key; API keys are for the separate Higgsfield API. MCP draws from the credits already in your plan - but it draws from them even when your plan says unlimited.
Is there a free trial for Higgsfield MCP?
No. Higgsfield's plans article says there is no trial period, and the connection guide requires an active paid subscription.
Why did my unlimited plan charge me credits?
Because that is the documented behavior. Unlimited applies on higgsfield.ai only, not through MCP, CLI, Canvas or Supercomputer.
Is there a rate limit?
Higgsfield does not publish one. The practical constraint is concurrency - Starter allows 2 parallel videos and 4 images, and Plus on monthly billing allows 6 and 8 (the annual Plus card lists unlimited paid parallel generations) - and a paid Concurrency Boost can add up to 16 more.
Can I use it in ChatGPT the same way as Claude?
Mostly, through the official plugin, but audio generation and website building are not available there.
Who is responsible for what an agent generates?
You are. Higgsfield applies its Developer Terms and states you are responsible for anything an agent generates through your account.
Who this is for
Higgsfield MCP is worth connecting if you already run agents daily and want generation inside that loop, if you value a bundled workflow over a large raw model catalog, and if you are on Plus or Ultra so the credit budget is not the first thing you hit.
Skip it if you are on Starter and expected unlimited output, if your plan was to get around credit limits by going through an agent, or if you only generate a few clips a month - the web app gives you strictly more for the same money.
The tool is good. The confusion around it is entirely about a pricing rule Higgsfield documents clearly and most write-ups skip. Now you know which side of it you are on.
Explore the available options
Confirm the final price, eligibility, and refund terms on the seller page before purchasing.
