10 Best AI Video MCP Servers for Claude, ChatGPT and Cursor in 2026

Ten MCP servers let an AI agent generate video from Claude, ChatGPT, Cursor or a coding agent: Higgsfield, Runway, HeyGen, fal.ai, Dora, Vivideo, AITuber, OpenArt, Imagine.art and the open-source mcp-video-gen. They differ in which clients they document, whether an API key is required, and whose balance the generation hits. All facts below were checked on each vendor's own pages in October 2026.

How We Compared the Servers

The list includes servers that publish their own MCP documentation for at least one of Claude, ChatGPT or Cursor, plus one open-source option for self-hosting. Every server was then checked on the same criteria, taken from the vendor's own MCP page or developer docs. Facts verified October 17, 2026.

• Clients. Which of Claude, ChatGPT, Cursor and Claude Code the vendor documents, and how: connector URL, plugin, marketplace or CLI.
• Models. One in-house model or a multi-model catalog.
• Setup and auth. OAuth sign-in in the browser or an API key pasted into a config file.
• Billing. Credits of an existing plan or pay-per-run.
• Long renders. Whether the server returns a job ID and a status tool, so a render that takes minutes does not die on a client timeout.
• Beyond generation. Characters, audio, ad workflows, publishing.

Where a vendor page does not state something, the card says so. Higgsfield takes the first slot because one connector covers a multi-model catalog together with characters, audio and ad workflows, all on one account and one plan-credit balance, with no separate API key to manage.

AI Video MCP Servers at a Glance

1. Higgsfield MCP: One Endpoint for Claude, ChatGPT and Cursor Without an API Key

Higgsfield MCP turns the models, characters and ad workflows of a Higgsfield account into tools an agent can call, billed to the credits that account already has. One endpoint, https://mcp.higgsfield.ai/mcp, works as a custom connector in Claude, an official plugin in ChatGPT and a marketplace add-on in Cursor. Coding agents go a different route: Claude Code, Codex, OpenClaw and Hermes install Higgsfield CLI with its companion skills and call it from there. You sign in with your Higgsfield account, so there is no key to create or rotate.

In our check of the MCP model catalog on October 16, 2026, the same connector covered characters, audio and ad workflows, so a Seedance clip, a Soul character and a Marketing Studio ad share one account and one credit balance. Setup steps for each client are in the Help Center article How do I connect Higgsfield to an AI agent?

What to keep in mind

• A paid subscription is required, and every generation through the connector uses credits.
• ChatGPT has two gaps: audio generation and the Website Building skill are not available there.
• No hard credit cap per session: spending limits are a prompt-level instruction to the agent.

Best for: creators who want video, characters, audio and ads from one connector inside the Higgsfield AI-native creative suite.

2. Runway MCP: A Plan-Based Connector With Plugins for ChatGPT and Cursor

Runway MCP connects Claude.ai, Claude Desktop, ChatGPT, Grok Bot, Cursor and Replit to a Runway account at https://mcp.runwayml.com/mcp. In ChatGPT, Grok Bot and Cursor it installs as a plugin. Elsewhere you paste the URL and sign in. No API key is needed.

Runway's own examples on the MCP page are product ads and marketing videos built from a product link. Runway Dev MCP is a separate server for API builders and does not generate media.

What to keep in mind

• Explore Mode is not available through MCP, so every generation uses credits.
• Polling is undocumented, which matters for renders that outlast a client timeout.
• Model access depends on the plan, as Runway's FAQ states.

Best for: Runway subscribers who want their existing plan available inside agents.

3. HeyGen Remote MCP: Avatars, a Video Agent and Batches of 100

HeyGen Remote MCP is a hosted endpoint at https://mcp.heygen.com/mcp/v1/. It is documented for Claude Web, Desktop and Code, Codex CLI, Cursor's marketplace, Gemini CLI, Lovable, Manus, Superhuman and OpenAI, and HeyGen's MCP landing page has an Add to ChatGPT button. Authentication is OAuth, and all HeyGen plans qualify.

The Video Agent scripts, casts and renders from one prompt. The scene model takes text, an image or up to twelve references.

What to keep in mind

• One vendor's models, so there is no multi-vendor catalog to switch between.
• Custom agents need whitelisting if they run on a domain HeyGen has not approved.
• The toolset is built around presenters, with one scene model for clips without an avatar.

Best for: teams producing presenter-led or translated videos at batch scale.

4. fal.ai MCP: Raw Model Access With Explicit Queues

fal documents two routes to its model-generation server. The per-client guides for Claude, Claude Code, Codex CLI and Cursor use the OAuth relay at https://mcp.fal.ai/mcp-relay: you sign in in the browser and paste no key, and ChatGPT gets the same server as a plugin. The MCP overview page still show the standard endpoint https://mcp.fal.ai/mcp with a FAL_KEY in the client config for Claude Code, Claude Desktop, Cursor and Windsurf. A separate genmedia CLI uses a local API key as well.

The docs are explicit about video: use submit_job then check_job from the start, and avoid resubmitting to check progress, because each resubmission is a new billable job.

What to keep in mind

• Outputs are raw, so characters, audio tracks and ad formats are assembled by you.
• Eleven tools cover discovery, pricing, queues and uploads; the agent picks among them.
• Team credits require selecting an Active MCP account in fal's settings.

Best for: developers who want the fal catalog inside an agent and control over the queue.

5. Dora: Sixteen Models With the Cost Shown Before Generating

Dora is a hosted server at https://doravideo.com/mcp for ChatGPT, Claude, Cursor and Google Antigravity. Authentication is OAuth 2.1 or a static API key. The server exposes four tools on purpose: list_models, generate_video, generate_image and check_job.

Dora shows cost before generation. list_models returns the coin cost of every model, sorted cheapest first, and coins are refunded automatically when a job fails. Dora's page states its figures were verified on September 5, 2026.

What to keep in mind

• No public REST API, so access is the web app, the iPhone app or MCP.
• No publishing or scheduling; Dora returns files you post yourself.
• Costs are defaults, so longer or higher-resolution clips cost more than list_models shows.

Best for: agents that need to compare and approve cost before each generation.

6. Vivideo MCP: A Hosted Server or an Open-Source Package, With Avatars and an Auto Pipeline

Vivideo’s developer docs lead with the hosted server: point Claude Desktop, Claude Code, Cursor or any MCP client at https://api.vivideo.ai/mcp with your API key in the Authorization header, and there is nothing to install. The same 13 tools also ship as an npm package, npx -y @vivideo/mcp, for anyone who wants to run the server locally with the key in the env block. The package is open source under MIT.

Thirteen tools are listed, from create_auto_video to list_models. Vivideo documents nine agent skills that teach an assistant to check credits first and avoid duplicate renders.

What to keep in mind

• An API key lives in the config file; there is no OAuth path.
• No ChatGPT setup guide is published; Vivideo’s ChatGPT presence is a separate GPT, not the MCP server.
• The catalog lists Seedance 2.0, with newer versions not yet in the API table.

Best for: builders who want an auditable open-source server with avatars and voices.

7. AITuber: Finished Narrated Videos and Standalone Clips Published From the Chat

AITuber is a hosted server at https://mcp.aituber.app. It is documented for Claude (Desktop, claude.ai and Claude Code with one connector), ChatGPT through a developer-mode custom connector, Cursor, Codex, Notion AI, OpenClaw and Hermes. You sign in with your account, or pass an API key as a header for unattended agents.

AITuber’s own FAQ draws the line this way: it returns a finished narrated video, while model servers such as fal.ai return raw clips. The server also has a create_clip tool for one standalone clip of 1 to 30 seconds without narration, and a clip-models endpoint that lists the available models with their costs. Videos run from 15-second Shorts to 20 minutes.

What to keep in mind

• Less model control in full videos: the pipeline handles script, voice, visuals, captions and assembly, so you do not pick a different model per shot.
• Standalone clips need a paid plan and are capped at 30 seconds each.
• ChatGPT needs developer mode for the custom connector.

Best for: channels that want a narrated, captioned video published from one prompt.

8. OpenArt MCP: Library and Projects Inside the Agent

OpenArt's connector at https://mcp.openart.ai/mcp is documented for Claude as a custom connector, ChatGPT as a plugin activated with @OpenArt, Cursor through mcp.json, Kimi as a plugin, and other agents over Streamable HTTP. Authentication is OAuth, with no API key.

For Claude Code and Codex, OpenArt itself points to its CLI with openart login.

What to keep in mind

• Character and Smart Shot are on the roadmap and are listed as unavailable through MCP for now.
• Managed workspaces on Claude Team or ChatGPT Business need an admin to add the connector first.
• Naming OpenArt in the prompt is recommended when several MCP servers are connected.

Best for: teams that already keep their assets and projects in OpenArt.

9. Imagine.art MCP: 96 Tools and Eleven Clients From Two Commands

Imagine.art MCP runs at https://mcp.imagine.art. Two commands, npm i -g @imagineartofficial/mcp and imagine install, set it up in eleven clients: Claude, Claude Code, Cursor, Codex, VS Code, Windsurf, Cline, Gemini CLI, Hermes, OpenClaw and Grok. ChatGPT uses a plugin from the directory. Claude on the web needs only the URL. Authentication is OAuth.

Seven tools cover the common requests. The other 89 are job-specific, from UGC avatars to 3D logo animation.

What to keep in mind

• Clips run five or ten seconds, depending on the model.
• 96 tools is a long list, so the agent may need steering toward the right one.
• Some pipelines need desktop apps, such as Photoshop or After Effects, installed locally.

Best for: commerce and fashion teams running creative pipelines through an agent.

10. mcp-video-gen: Your Own Weights Behind Your Own Endpoint

mcp-video-gen is an Apache-2.0 project on GitHub for people who want to run their own models. It is a self-hosted server, deployed through Portainer from a single video-mcp.yml stack, and it can expose a local ComfyUI instance and, optionally, a host-installed Blender.

You publish it behind your own endpoint, typically through a Cloudflare Tunnel with Cloudflare Access verification. The README documents native file upload from ChatGPT.

What to keep in mind

• No built-in generation workflows, memory or skills; quality is your ComfyUI setup.
• A single contributor maintains the project, with the last release two months before this check.
• A GPU host you administer is a prerequisite.

Best for: developers who want video generation on their own infrastructure.

How MCP Servers Handle Long Renders

A video render usually takes longer than a single MCP call will wait, so the job model matters. The documented pattern is a job ID plus a status tool.

On fal, resubmitting a generation to check progress creates a second billable job, so use the existing request ID and the status tool instead. On every other server, the documented status or wait tool is the way to follow a render, not a second generate call.

Which Server Fits Which Client

The table below groups the same ten servers by the client you already use. Only setups the vendor documents are listed.

For Higgsfield, the documented route depends on the client: the MCP connector in Claude web and Desktop, the official plugin in ChatGPT, the Marketplace integration in Cursor, and CLI + Skills for Claude Code, Codex, OpenClaw and Hermes.

What to pick comes down to the workflow. In Claude, Higgsfield covers multi-model creative production with characters and audio on one balance, HeyGen fits presenter and localization work, and fal.ai gives raw model access with an explicit queue. In ChatGPT, take the vendor’s official plugin where one exists; a pasted server URL in developer mode is the fallback for Dora and AITuber. In Cursor and Claude Code, fal.ai gives the most control over models and queues, Vivideo works through a hosted URL or a local package with an API key, and Higgsfield brings its full catalog through the Marketplace in Cursor and CLI + Skills in Claude Code. For your own weights, mcp-video-gen is the self-hosted option.

Who This Is For

Creators and developers who already work in Claude, ChatGPT, Cursor or a coding agent and want video as a tool call. With Higgsfield MCP, a clip generated from Claude shows up in Assets next to the ones made on higgsfield.ai, can reuse a Soul character trained there, and is paid for from one balance. Step-by-step guides exist for Claude and ChatGPT, and the full client list is on the Higgsfield MCP page.