The Best AI Image and Video Generation MCP Servers for Claude and ChatGPT
ClipSpeedAI
◷ 11 min read
Updated
For most creators, the AI generation MCP servers worth connecting in October 2026 are Higgsfield (its whole model menu on paid-plan credits), Runway (Gen-4.5 plus Seedance 2.5, Kling and Nano Banana Pro on its Pro and Max plans) and fal (1,000+ models and a tool that reads the price before a run). Replicate suits developers who already pay per prediction, and HeyGen is the pick for talking presenters. None of them enforces a per-chat spend cap for you, so the brakes are your balance and your approvals.
We read each provider's own MCP page, help center and pricing on October 8, 2026 and ran no generations, so every capability below is as documented. New to the protocol? Start with what MCP is.
Which AI generation MCP servers are worth connecting?
A generation MCP server lets Claude or ChatGPT write the prompt, set the parameters and start a render on the provider's GPUs, billed to your account there. These are the seven we could check against each provider's own documentation.
| Server | How you connect | What it makes | Price before a run | Spend ceiling | Where results land |
|---|---|---|---|---|---|
| Higgsfield | URL in Claude, ChatGPT plugin; sign-in, paid plan | Every model on its menu, plus audio and a YouTube clipper | Ask the agent; rates on its pricing page | No native cap | Your Higgsfield Assets |
| Runway | URL in Claude, ChatGPT plugin; sign-in, Pro plan or above | Gen-4.5, Seedance 2.5, Kling, Veo, GPT Image 2, Nano Banana Pro | Per-second credit rates published | Your credit balance | The chat and your Runway library |
| fal | Hosted relay with OAuth, ChatGPT plugin | 1,000+ models | Yes, a pricing tool | Prepaid balance | File links in the chat |
| Replicate | Remote server or local package; API token at sign-in | Public Replicate models | Per-model prices on its site | Prepaid balance | Links, removed after an hour by default |
| HeyGen | URL, ChatGPT plugin; OAuth | Avatars, Video Agent, voices, translation | Not on its MCP page | Premium credits in your plan | The same thread |
| Google Genmedia | Self-hosted; Google Cloud credentials | Veo, Gemini Omni, Nano Banana, Lyria, speech | Your Cloud pricing | Your Cloud billing | Your storage bucket |
| Hugging Face | Hosted; setup from your account's MCP settings | Community Gradio Spaces | Varies by Space | Not documented | Images in the chat |
Checked on each provider's own pages on October 8, 2026. Monthly list prices without promotions.
Which one fits your use case?
- Many models on one subscription: Higgsfield, with per-call approval on.
- Pricing a shot from its length: Runway.
- A price check before every run: fal.
- A developer already billing through Replicate: Replicate.
- Presenters, digital twins and translation: HeyGen.
- Your own Google Cloud project and quotas: Google Genmedia.
- Experiments with open models: Hugging Face.
Generated the shot? The long videos still need clipping.
ClipSpeedAI turns your streams, podcasts and long uploads into vertical clips with burned-in captions and a viral score for each, from the web or from Claude.
Try ClipSpeedAI →Higgsfield MCP: every model, one subscription, no brakes
What it offers. Higgsfield's server at mcp.higgsfield.ai/mcp works as a Claude custom connector, a ChatGPT plugin and a Cursor marketplace add-on, and terminal agents such as Claude Code install its CLI. Its help center says an active paid subscription is required and no API key is needed. Over MCP you get image and video generation on all its models, upscaling, Soul characters, audio such as voiceovers and dubbing, and a Personal Clipper that cuts YouTube videos into shorts. Results land in your Higgsfield Assets.
Pricing. Starter is $19 a month for 270 credits, Plus $59 for 1,200 and Ultra $129 for 3,000, billed monthly. The pricing page lists per-model rates, such as about 22 credits per five seconds of Seedance 2.0 at 720p and 2 credits per Nano Banana Pro image.
Every agent generation spends credits: the unlimited models and free generations on its plans work only on higgsfield.ai, not over MCP. The help center says there is no native credit cap and suggests a spending instruction in the chat, which it calls a prompt-level instruction, not a hard limit. Audio generation isn't available in ChatGPT.
Runway MCP: credit math you can do in your head
What it offers. Runway announced its hosted server on May 27, 2026, at mcp.runwayml.com/mcp. In ChatGPT, Grok and Cursor you add the Runway plugin and sign in; Claude connects by URL, with no API key. Which models the agent can use depends on your plan: Gen-4.5, Seedance 2.5, Kling, Veo, GPT Image 2 and Nano Banana Pro among them. Generations come back to the chat and are saved in your Runway library.
Pricing. Gen-4.5 uses 12 credits per second, so a 10-second clip is 120 credits. The pricing page lists Runway MCP under Pro and Max but not Standard; Pro is $35 a month billed monthly, with 2,250 credits.
No spend control is described beyond your balance, Standard and Pro credits reset each billing cycle, and the MCP page doesn't mention posting. Runway archived its older open-source API server on GitHub at the end of September 2026, so use the hosted one.
fal MCP: the widest catalog, with a price check built in
What it offers. fal's relay at mcp.fal.ai/mcp-relay uses browser OAuth from Claude, Claude Code, Codex CLI and Cursor, and ChatGPT reaches it through fal's plugin. Eleven tools cover search, pricing, generation, job management and uploads, over a catalog fal's launch post puts at 1,000+ models.
Pricing. get_pricing is the standout: the assistant can read a model's price before it runs anything. Billing is prepaid credit, video models charge per second or per video, and fal says server errors are never billed.
fal warns that calling run_model or submit_job again to check progress creates a new billable job, so long video renders should go submit_job, then check_job, then get_job_result. Its docs say the example prompts add no server-side approval gate. Outputs are file links, not a library, and with a catalog this wide you should name the model you want.
Replicate MCP: for developers who already pay per prediction
What it offers. Replicate's remote server at mcp.replicate.com is the route for claude.ai; Claude Desktop, Cursor and VS Code can also run the local replicate-mcp package. Sign-in is a web flow where you provide a Replicate API token. The tools cover the operations in Replicate's HTTP API, such as searching models and creating predictions, and an experimental code mode swaps them for two tools that search the SDK docs and run TypeScript in a sandbox.
Prepaid credit lasts a year, and the optional auto-reload adds money rather than capping it. New customers can't set spend limits since July 1, 2025. Output files from API predictions are removed after an hour by default, so download what you keep. ChatGPT isn't among its documented clients.
HeyGen MCP: presenters, twins and translation
What it offers. HeyGen's hosted server at mcp.heygen.com/mcp/v1/ uses OAuth with no API key, and works as a ChatGPT plugin and in Claude, Claude Code, Codex, Cursor, Gemini CLI and others. It reaches Video Agent, stock avatars and your digital twin, voices, translation with lip-sync and clipping of long recordings.
Pricing. Renders draw on the premium credits in your plan, with no separate API charge. Creator is $29 a month billed monthly, with 600 credits.
The MCP page lists no credit cost per render and says nothing about posting. A digital twin renders only with that person's recorded consent on file. It's a presenter tool, not a general b-roll or product-shot generator.
Are there self-hosted or open options?
Google Genmedia
Google Cloud's genmedia-creative-studio repository ships Go MCP servers for Gemini image models (Nano Banana), Veo, Gemini Omni, Gemini text-to-speech, Chirp 3 HD voices, Lyria music and an audio-video compositing tool. They sign in with your Google Cloud credentials, so cost follows your Cloud project, and files go to a storage bucket you choose.
The repository says it is not an officially supported Google product, and setup assumes a Cloud project and comfort on the command line.
Hugging Face
Hugging Face's server connects Codex, Cursor, VS Code, Zed, ChatGPT and Claude Desktop to the Hub. Image generation runs through community Gradio Spaces that you add in your MCP settings, or that the assistant finds at runtime if you turn on Dynamic Spaces.
Quality, speed and uptime depend on whichever Space you call. Fine for experiments, not for client work.
Where does ClipSpeed fit?
ClipSpeed runs two MCP servers, and they cover different halves of the job.
ClipSpeed Create (early access) at https://api.clipspeed.ai/mcp/create brings ClipSpeed's AI Creator lineup into Claude and ChatGPT: Seedance 2.5, Kling 3.0 Turbo, MiniMax H3 Max, Veo 3.1 and Gemini Omni 1.1 Flash for video, and Nano Banana Pro, GPT Image 2.5 Flare and GPT Image 2.5 Sunburst for images. Every generation is quoted first and runs only after the account owner approves the quote on a signed-in ClipSpeed page, outside the chat. None of the other servers above documents an approval step on the server side.
The clipping connector at https://api.clipspeed.ai/mcp works in Claude today. It cuts a long video or a live Twitch, Kick or YouTube stream into captioned 9:16 shorts ranked by viral score, with six caption styles, and returns a title, an opening hook and a suggested posting time for each clip. It can publish a finished clip to your connected YouTube channel after you confirm the clip and the destination. Setup is on the ClipSpeed MCP page, and the Claude guide walks through a full session.
How do you stop an agent from burning your credits?
Every generation call spends money, and the protections are thin. Claude's help center warns that custom connectors aren't verified by Anthropic, that a malicious server can hide instructions, and that "Allow always" belongs only on tools you trust to run unsupervised. Research mode can call connector tools without further approval. In ChatGPT, full MCP with write actions is a beta for Business, Enterprise and Edu, while Pro accounts can connect MCP servers with read and fetch permissions only, which is why most generators reach ChatGPT as plugins. Our MCP security notes go further.
- Fund small. Keep balances near a week of use, with auto-reload off.
- Keep approvals on. Ask for approval per call on generation tools, and set tools you don't need to Blocked under Customize > Connectors.
- Set a standing rule. Paste the block below at the start of every generation chat.
- One generator per chat. Two connected generators invite the agent to try both.
- Save outputs yourself. Links can expire.
- Disconnect when done. Remove servers you no longer use.
Standing spend rule:
Before you call any tool that generates an image, video or audio:
1. Tell me the model, duration, resolution, aspect ratio and number of outputs.
2. Tell me the exact cost in credits or dollars, and my remaining balance if you can read it.
3. Wait for me to reply "approve". Silence or "ok" is not approval.
Never retry a failed or rejected generation without asking.
Never switch to a more expensive model or a higher resolution on your own.
Stop and ask if this chat's total would pass 100 credits or $5.To set those thresholds, see what AI video generation actually costs.
What does a generate-then-finish workflow look like in Claude?
- Brief with the full spec. Paste a prompt like the one below and ask for a quote.
- Approve one take. Review it before asking for variants.
- Refine by naming what stays. "Keep the subject, light and framing; slow the push-in." One change per run.
- Edit it in. Use the shot as a hook or cutaway in your longer video.
- Finish with clipping. Ask Claude to clip the finished video into captioned shorts, then publish the best one.
Video prompt with a cost check:
SCENE: 8-second single take, vertical 9:16, real time. A potter opens a top-loading
kiln and lifts out a glossy teal vase.
SUBJECT: Woman, early 40s, cropped grey hair, rust canvas apron over a white T-shirt,
tan suede kiln gloves. Same face and outfit throughout.
LOCATION: Converted garage studio, pegboard of clay tools, shelves of unglazed pots.
CAMERA: Chest height, about 1.2 m from the kiln. One slow push-in across the full
8 seconds. No cuts, no zoom.
ACTION: 0-2 s she lifts the lid with both hands. 2-5 s she grips the vase at its neck
and base. 5-8 s she raises it to eye level and turns it a quarter turn, then pauses.
LIGHTING: Window key from camera left, late afternoon. Warm kiln glow from below.
AUDIO: Kiln fan hum, a soft ceramic clink, and one line from her: "Still warm." No music.
STYLE: Photoreal, natural grade, light film grain. No text, no logos, no watermarks.
Before generating, tell me the model, resolution and exact cost, then wait for my approval.Our guide to generating AI video from Claude or ChatGPT covers connector setup step by step, and the AI video prompt guide explains each block.
How should you choose? A checklist
- Sign-in: revocable OAuth beats a key pasted into a config file.
- Cost before a run: a pricing tool or published per-second rates.
- A ceiling you control: a prepaid balance or plan credits; no third-party server here enforces a per-chat cap for you.
- Approvals that work in your client: per-call approval in Claude, plugin permissions in ChatGPT.
- File storage: a library, a chat link or an expiring link.
- Consent: written permission for any real face or cloned voice you feed in.
Once the shots are made, the long-form footage around them still has to become clips. Try ClipSpeedAI's clipping for that half.
Sources and method
We read each provider's MCP page, help center, documentation and pricing page on October 8, 2026, with pricing toggled to monthly billing. We ran no generations; where a provider doesn't publish a detail, such as a per-render cost or a spend cap, the table says so instead of guessing. For a broader list of creator MCP servers, see the best MCP servers for creators and MCP servers for video workflows.
- Higgsfield: connect guide, MCP, pricing. Runway: launch, MCP, pricing
- fal: MCP docs, pricing, launch post. Replicate: MCP, credit, retention, spend limits
- HeyGen: MCP, pricing. Google Genmedia, Hugging Face MCP
- Clients: Claude custom connectors, adding a connector, ChatGPT developer mode
Got any questions left?
Can Claude generate images or videos by itself?
No. Claude renders no media itself; it calls a generation server through a connector, and that provider renders and bills on your account. ChatGPT makes still images with ChatGPT Images but relies on plugins such as Runway's or Higgsfield's for video.
Which MCP server is cheapest for AI video?
It depends on the model more than the server. Pay-as-you-go balances on fal or Replicate suit occasional use; plan credits suit steady volume. On Runway, a 10-second Gen-4.5 clip is 120 credits, and fal's get_pricing tool reads any model's price before a run.
Do I need an API key to connect these servers?
Usually not. Higgsfield, Runway, fal and HeyGen use account sign-in. Replicate asks for an API token during its web sign-in, and Google Genmedia uses your Google Cloud credentials.
Can I use these servers in ChatGPT?
Higgsfield, Runway, fal and HeyGen offer ChatGPT plugins. Adding another server by URL needs developer mode, and OpenAI currently limits full MCP with write actions to Business, Enterprise and Edu workspaces; Pro accounts get read and fetch permissions only.
Is it safe to set a generation connector to "Allow always"?
Only if you accept spending you didn't approve. None of the third-party servers here documents a per-chat cap, so per-call approval plus a standing cost rule is the safer default. Claude's help center says to use "Allow always" only for tools you trust to run unsupervised.
How does ClipSpeed Create handle spending?
ClipSpeed Create is in early access. It quotes every generation first, and the generation runs only after the account owner approves that quote on a signed-in ClipSpeed page, outside the chat. Its lineup is Seedance 2.5, Kling 3.0 Turbo, MiniMax H3 Max, Veo 3.1, Gemini Omni 1.1 Flash, Nano Banana Pro, GPT Image 2.5 Flare and GPT Image 2.5 Sunburst.
Related guides
Published by ClipSpeedAI · AI video generation and AI clipping in one place — create with Seedance, Veo, Kling and Nano Banana, then cut it into captioned shorts.