CapCut MCP: The Official ChatGPT Plugin vs the Community Servers

CapCut now has an official way for an AI assistant to edit video: a ChatGPT plugin, announced at CapCut World in Los Angeles on 24 October 2026. It is not a general MCP server, though. It runs inside the ChatGPT desktop app, builds an editable draft, and hands you to CapCut's editor to finish. If you want CapCut from Claude, Cursor or a script, the community servers that write CapCut desktop drafts on your own computer are still the only route. This guide covers both, shows how to set up a community server, and covers the cloud option for when you need a finished video file. Wireflow runs that cloud side, with generation, timeline edits and rendering behind one hosted MCP server that works in any client that takes a remote MCP URL.

Quick answer

• CapCut announced an official ChatGPT plugin on 24 October 2026. CapCut's page calls it CapCut x Codex and lists the ChatGPT desktop app on PC.
• It is not a standalone MCP server. CapCut publishes no MCP endpoint and no public editing API, so Claude, Cursor and other clients cannot connect to it.
• The plugin builds an editable draft in chat, then you refine and export in CapCut's editor.
• Community servers (VectCutAPI and its wrappers, plus standalone projects) still write local CapCut desktop drafts for any client that runs stdio servers.
• A hosted MCP server edits and renders in the cloud and returns a video file, from any MCP client.

What CapCut's official ChatGPT plugin does

CapCut announced the plugin on stage at its first CapCut World event, and its post about it passed a million views on X within two days. The flow it describes is short: start with an idea in ChatGPT, let the CapCut plugin build a draft video in the chat, then open that draft in CapCut's editor to refine, share and publish.

CapCut's product page fills in a few details:

• Client: the ChatGPT desktop app on PC. CapCut says it is "not a browser-only" tool and that availability varies by app version, account and region. Creator posts at the launch say mobile and web support should follow in mid-November; CapCut's own page does not give a date.
• Install: you copy a prompt from CapCut's page into a ChatGPT task and follow the authorization steps.
• What it edits: prompts, scripts, images or clips go in. It builds a smart rough cut, can show the timeline and its tracks in chat, and takes chat edits to sequence, duration, media, text, audio, captions, transitions and effects. It also searches CapCut's templates and stock library.
• Output: an editable draft. CapCut calls generation "a starting point for editing", and export happens in CapCut.
• Price: CapCut says you can start for free. It lists no paid tiers or usage caps for the plugin yet.

What it is not: an MCP server you can add to another client. ChatGPT's plugin system is built on MCP, but CapCut has not published an endpoint, and its own FAQ says "CapCut MCP" is not how it describes the product. Nothing announced so far covers Claude, Cursor, Codex CLI outside ChatGPT, or scheduled jobs with no one at the keyboard. Reviewers at launch asked the same question about Claude.

	CapCut ChatGPT plugin (official)	CapCut MCP (community)	Hosted MCP server (Wireflow)
Who makes it	CapCut	Open-source authors	Wireflow
Where you use it	ChatGPT desktop on PC (mobile and web reported for November)	Claude Code, Claude Desktop, Cursor, any stdio client	Claude, Claude Code, Codex, Cursor, ChatGPT (developer mode), Grok, any remote MCP client
Where it runs	CapCut's cloud	Your computer, next to CapCut desktop	Remote, over Streamable HTTP
Output	Editable draft, finished in CapCut	Editable CapCut desktop draft	Rendered video file, up to 180 s at 1080p and 60fps
Who exports	You, in CapCut	You, in CapCut desktop	The agent, with render_video
New footage	CapCut's generation tools and stock library	Bring your own clips	Image, video and audio models on the same board
Agent checks its own work	Timeline shown in chat	No	Free preview stills and check_video playback checks
Runs with no one there	No	No	Opt-in automations and webhook triggers
Hand editing	The full CapCut editor	The full CapCut editor	Scene-graph edits plus the Wireflow video editor
Price	Free to start, tiers not announced	Free and open source	Credits for generation; reads, edits, previews and renders spend none at current pricing

Where CapCut wins: the editor itself, its template and asset library, and a huge user base that already knows the app. If a person finishes every cut by hand, the official plugin is the easiest start and the price is right. Where the hosted route wins: the agent gets a finished file back, can check its own frames, works from whichever assistant your team already uses, and can run on a schedule.

How a CapCut MCP server works

MCP (Model Context Protocol) lets an AI assistant call tools that a server exposes, such as create_draft or add_text, and read the results back into the chat. You type "make a 30 second vertical cut from these five clips with a title card", and the assistant turns that into a sequence of tool calls. It is the same pattern behind the best MCP servers for video editing, where Claude calls trimming, captioning and rendering tools through one server.

CapCut itself still has no server you can point a client at. Its official route is the ChatGPT plugin above, which lives inside ChatGPT. It publishes no developer API for editing either; our look at the CapCut API question found no developer portal on its domains. So for any other assistant, the community servers are the way in.

So the community servers work on files instead. CapCut desktop stores each project as a folder of JSON files, including draft_content.json and draft_meta_info.json, inside a drafts directory. A CapCut MCP server writes or edits those files, and you open the project in CapCut to preview, adjust and export. Nothing touches your CapCut account or cloud projects, which is the same local-first approach some teams take when they connect Claude to video editing apps on their own machine.

The main CapCut MCP servers compared

Five projects come up most often in MCP directories. They fall into two families: wrappers around the VectCutAPI backend, and standalone servers that write draft JSON directly. Unlike hosted video editing APIs, none of them renders anything on a server.

Server	How it works	Tools	Needs	Main caveat
VectCutAPI (sun-guannan)	Python HTTP server on port 9001, plus an mcp_server.py stdio mode	Drafts, video, audio, text, subtitles, effects, stickers, keyframes	Python 3.10+, FFmpeg, CapCut or Jianying desktop	Saves a dfd_ folder you copy into the drafts directory; its cloud rendering is not open source
Atx-Guy/capcut-mcp-server	TypeScript MCP wrapper over VectCutAPI	11 tools, from capcut_create_draft to capcut_save_draft	Node.js plus a running VectCutAPI backend	npm package not yet published, so you build from source
alex28042/capcut-mcp	Writes draft JSON directly	7 tools: create_draft, list_drafts, get_draft, add_video, add_image, add_audio, add_text	Node.js, pnpm, CapCut desktop, ffmpeg recommended	No effects, transitions, keyframes or export; validated on CapCut 3.3.0 for macOS
JmsLdrn/capcut-mcp	Edits drafts by copying segment templates from a known-good draft	17 tools, including move, trim, split and delete segment	Node 18+, ffmpeg and ffprobe, CapCut desktop	Effects and transitions are best effort; assumes a Windows layout
dlyrr/capcut-mcp	Builds new projects from clips and can cut on the beat	6 tools, including capcut_detect_beats and capcut_create_beat_edit	Node 18+, ffmpeg and ffprobe, one existing CapCut project	Author calls it "extremely experimental"; tested against the file format, not CapCut itself

The VectCutAPI family covers the most ground, including subtitles, effects and keyframes, because its backend models more of the draft format. A fork, MigueDuque/capcut-mcp-server-extended, adds font and position controls plus talking-head and Reels/Shorts presets for Claude Code. The standalone writers are lighter to install but cover less: alex28042's server edits four segment types, while JmsLdrn's adds trimming, splitting and a save step that refuses to write while CapCut is open. If the timeline you want is mostly clips plus captions, a headless video editor built for automation is another way to reach it.

How to set one up with Claude Code

Setup differs by family. The standalone writers need only Node, while the VectCutAPI family needs its Python backend running first. Both register as local stdio servers, a different model from the hosted AI video generation MCP setup, where you paste a URL.

1. Install CapCut desktop and open it once so the drafts folder exists. Close it before the agent writes, because CapCut's autosave can overwrite changes.
2. Install ffmpeg so the server can read clip durations and dimensions.
3. Clone and build the server (npm install and npm run build, or pnpm install and pnpm build).
4. For the VectCutAPI family, start the backend with python capcut_server.py. It listens on http://localhost:9001.
5. Register the server with your client (commands below).
6. Ask for an edit, then open the draft in CapCut, check it, and export.

For a standalone writer in Claude Code:

claude mcp add --scope user capcut -- node /absolute/path/to/capcut-mcp/dist/index.js

For the VectCutAPI wrapper in Claude Desktop, add it to claude_desktop_config.json (on Windows, under %APPDATA%\Claude\):

{
  "mcpServers": {
    "capcut": {
      "command": "node",
      "args": ["/absolute/path/to/capcut-mcp-server/dist/index.js"],
      "env": { "CAPCUT_API_URL": "http://localhost:9001" }
    }
  }
}

A good first prompt is concrete: "Create a 1080x1920 draft, add clip_01.mp4 to clip_04.mp4 end to end, put the title 'Day 1 recap' over the first three seconds, and save." The assistant calls the create, add and save tools in order, and the project appears in CapCut's list. Keep media on a stable path such as ~/Movies, since drafts point to files by absolute path, and treat each prompt as one step in a larger LLM video editing workflow rather than a finished edit.

Where desktop draft files hit their limits

Most guides list the servers but skip what they cannot do. These limits apply to every project above:

• No video comes out. Every server stops at a draft. Someone opens CapCut and exports, so the agent cannot hand back an MP4 or post it anywhere.
• Local only. The servers run as stdio processes next to CapCut. Web connectors that expect a remote MCP URL, and agents running in the cloud, cannot start them.
• CapCut must be closed. CapCut keeps open projects in memory and can overwrite the file when it saves or closes.
• The format is undocumented. Draft JSON changes between CapCut versions and platforms, so an app update can break a server overnight. alex28042's server was validated on one version, and dlyrr's on drafts from a single Windows install.
• Paths are absolute. Drafts reference media by full path, so moving a clip or opening the project on another computer breaks the links.
• No triggers or schedules. Nothing runs unless someone is at the machine with the assistant open.
• Effects are partial. Transitions, animations and rich text styling are missing or best effort in most servers.

For agencies, these limits matter more than tool counts: a pipeline that needs a person to export each cut stalls after a few videos a day, the bottleneck covered in our guide to outsourcing video editing to AI.

The cloud option: a hosted MCP server that renders

A hosted MCP server moves the editing and the render off your machine. You add one URL as a connector, sign in with OAuth, and the agent calls tools on a remote backend that returns a finished file. Our hosted MCP server works this way over Streamable HTTP at https://www.wireflow.ai/api/mcp, with no API key to paste.

In Claude Code it is one command, the same shape as the fal MCP setup minus the key header:

claude mcp add --transport http video-mcp https://www.wireflow.ai/api/mcp

Its editor tools cover the jobs CapCut MCP servers do on drafts. get_video_timeline reads a Video Editor node's scenes and clips, edit_video_timeline replaces the scene graph, and preview_video_frames renders up to five still frames so the agent can check the cut. Reading, editing and previewing spend no credits, and render_video then renders the node in the cloud. The MCP editor reference lists the limits: renders up to 180 seconds at up to 60fps and 1920x1080, with 2 renders a minute and 30 a day per account.

Because the server is remote, the same job can run without anyone at a desk. Scheduled automations are an opt-in permission, and a workflow can also start from a plain HTTP call to its webhook trigger with no API key, the pattern behind a Zapier video editing integration.

	CapCut MCP (community)	Hosted MCP server
Where it runs	Your computer, next to CapCut	Remote, over Streamable HTTP
Setup	Clone, build, ffmpeg, sometimes a Python backend	One connector URL and an OAuth sign-in
Output	Editable CapCut draft	Rendered video file
Export	Manual, in CapCut desktop	render_video in the cloud
New footage	Bring your own clips	Image, video and audio models in the same server
Schedules and triggers	None	Opt-in automations and webhook triggers
Hand editing afterward	The full CapCut editor	Scene-graph edits and preview frames
Cost	Free and open source	Credits for runs; reads, edits and previews spend none

The trade-off is control. CapCut's editor is still better for frame-level hand work, keyframed effects and templates, so if a person finishes every cut anyway, a draft server is a fine way to skip assembly. If the agent has to finish the job alone, a video editing agent that renders server-side removes the desktop step.

Which setup fits which job

• You live in ChatGPT and finish by hand in CapCut: the official CapCut plugin.
• Rough-cut existing footage, then finish by hand: JmsLdrn/capcut-mcp, for its trim, split and backup-on-save.
• Beat-synced montages: dlyrr/capcut-mcp, accepting its experimental status.
• Drafts with subtitles, stickers and keyframes: VectCutAPI with the Atx-Guy wrapper.
• Quick clips-plus-title drafts on a Mac: alex28042/capcut-mcp.
• Finished MP4s, schedules, or agents that are not on your desk: a hosted MCP server.

Whichever route you pick, test it on a throwaway project first, pin your CapCut version, and if you go hosted, compare your volume against plan pricing.

Try it yourself: Build this workflow in Wireflow. It turns a prompt into a first frame, then a Video Editor node adds a slow zoom and a caption and renders a vertical MP4, the kind of finished file a CapCut draft cannot return.

FAQ

Is there an official CapCut MCP server?

Not as a server you can connect to. On 24 October 2026 CapCut announced an official plugin for ChatGPT, and ChatGPT plugins are built on MCP, but CapCut has not published an MCP endpoint for Claude, Cursor or other clients. The community servers are the only CapCut option outside ChatGPT, which is why most MCP tools for AI video generation come from model providers or hosted platforms instead.

Does the CapCut ChatGPT plugin work with Claude?

No. As of 26 October 2026 CapCut lists only the ChatGPT desktop app. To drive CapCut from Claude you need a community server and CapCut desktop on the same machine, or a hosted MCP server that does the editing and rendering itself.

Is the CapCut ChatGPT plugin free?

CapCut says you can start for free. It has not published paid tiers or usage limits for the plugin, and CapCut Pro and ChatGPT plan rules may still apply. Check CapCut's page before you plan volume around it.

Can the CapCut ChatGPT plugin export an MP4?

It builds a draft video in the chat and sends you to CapCut's editor to refine and export. If an agent has to deliver the file on its own, a video assembly API returns the render.

Can a CapCut MCP server export an MP4?

No. The servers produce an editable draft, and a person exports it from CapCut desktop. If you need the file itself from a single call, a video assembly API joins clips, captions and audio and returns the render.

Does CapCut MCP work with Claude Code and Claude Desktop?

Yes. Both can launch local stdio servers, using claude mcp add or claude_desktop_config.json. Connectors in the claude.ai web app expect a remote URL, so they cannot start a local CapCut server; an API with Claude Code integration is the remote route.

Is CapCut MCP free?

The servers are free and open source under MIT or Apache-2.0 licenses, and CapCut's own plans still apply to anything you do in the app. Hosted options charge per run instead, as described in our page on usage-based AI API pricing. CapCut's official ChatGPT plugin is free to start.

Why won't my generated draft open, or why is media missing?

Usually the CapCut version's draft format differs from what the server expects, media moved after the draft was written, or CapCut was open during the save. Close CapCut and keep media on a fixed path. An agentic video editing platform avoids the file-format problem by editing a timeline it controls.

Can I use CapCut MCP with Cursor or ChatGPT?

In ChatGPT, use CapCut's official plugin. Cursor and other clients that launch local stdio servers can run the community servers with the same command and path. Clients that only accept a remote HTTPS URL cannot. For coding agents that need a remote endpoint, see this video API for coding agents.

What is VectCutAPI?

VectCutAPI is an open-source Python project (Apache-2.0) that builds CapCut and Jianying drafts over HTTP on port 9001 and includes an MCP mode. Because it speaks HTTP, it can also sit behind an automation tool, the choice covered in n8n vs MCP for video automation.

Can an AI agent edit video without CapCut at all?

Yes. An agent connected to a hosted editor reads the timeline, changes scenes, previews frames and renders, all through tool calls. This agent that edits video page shows that setup end to end.

Conclusion

CapCut MCP now means two things. In ChatGPT, CapCut's official plugin builds a draft you finish in the CapCut editor, which is a good fit when a person does the last pass anyway. Everywhere else it still means community draft writers, limited by a local-only setup, an undocumented file format and a manual export. When the agent needs to deliver a finished video on its own, from whichever assistant you use, Wireflow's hosted MCP server handles generation, timeline edits and the render in one place, and our Claude video generation guide walks through that flow from a Claude chat.