Automating shorts with ChatGPT or Claude (MCP)
Yes, you can drive video clipping from ChatGPT or Claude through an MCP server: the assistant imports the video, waits for the transcript, picks the moments, reframes, captions and exports in 9:16 from a single sentence. As of October 2026, Opus Clip (27 tools) and ClipOnIt (25 tools, from the Pro plan at $49/month) offer such a server; OpenClip, OpenShorts and ChatCut do too. What does not automate: editorial judgment on which moments to keep, and publishing, which remains manual on several of these tools.
What is an MCP server, in one sentence?
The Model Context Protocol (MCP) is an open standard that lets an AI assistant (Claude, ChatGPT, Cursor...) call a third-party application's functions as if they were its own tools. The application exposes a list of named tools ("import a video", "export a clip") with their parameters; the assistant discovers them on its own and chains them to answer a plain-language request. For the user, it replaces a click-through interface with a conversation: "import this YouTube video, keep the three best moments, karaoke captions, export in 9:16 and give me the links".
Why automate shorts with an assistant rather than a script?
An API script always does the same thing. An assistant adapts the flow to each video: it can read the transcript, drop an off-topic moment, ask for confirmation before exporting, or summarize the produced clips in a message. The three cases where it pays off:
- Volume: more than two episodes a week, or several channels to process.
- Agencies: the same flow (import, selection, the brand's caption style, batch export) repeated for every client.
- Paid clippers: one VOD a day to cut fast; the economics are in how much do clippers make.
Below that threshold, a clipping tool's web interface is enough. Do ten episodes by hand first to learn which moments work; automate afterwards.
Which clipping tools have an MCP server in 2026?
| Tool | MCP server | Announced clients | Required plan | Source |
|---|---|---|---|---|
| Opus Clip | Yes, 27 tools (clip, detection, captions, reframing, filler-word removal, scheduled publishing) | Claude (Desktop, web, Code), Cursor, ChatGPT via Apps SDK | Free trial, then Pro metered per rendered minute | opus.pro/mcp, Oct. 2026 |
| ClipOnIt | Yes, 25 tools (import, wait, find a moment, bounds, reframe, auto-zoom, silences, captions, single or batch export, frame preview, quota) | Claude Code, Claude Desktop / claude.ai, ChatGPT (GPT Action or remote MCP) | Pro ($49/month) or Business | cliponit.app/.well-known/mcp.json, Oct. 2026 |
| OpenClip | Yes | Claude, ChatGPT | Not verified | openclip.app/mcp |
| OpenShorts | Yes (scored, captioned, 9:16 clips) | Claude, Cursor, ChatGPT | Not verified | mcpservers.org |
| ChatCut | Yes (silences, captions, reframing, B-roll) | ChatGPT | Not verified | chatcut.io |
Submagic, Vizard and Klap did not announce an MCP server on their public pages as of October 3, 2026, but expose a conventional API. The landscape moves fast; check each tool's page.
How it works in practice: a full flow
Example with the tools of ClipOnIt's MCP server, whose names are public; the other servers follow the same logic under different names.
- Connection. Create an API key in the app (Settings → API), then add the server to your client. With Claude Code: claude mcp add --transport http cliponit https://cliponit.app/mcp --header "Authorization: Bearer ck_…". With Claude Desktop or claude.ai: a custom connector on the same URL. With ChatGPT: a GPT Action from the OpenAPI file, or a remote MCP through the Responses API.
- Import. The assistant calls import_video_url with the YouTube link (or upload_video_file for a local file), then wait_video until transcription and moment detection finish.
- Selection. It reads the clip list with scores out of 100 and reasons. You can ask "keep the three highest-scored that talk about pricing"; the assistant filters on the transcript. For a specific passage, find_moment locates a plain-language description.
- Adjustments. set_clip_bounds then snap_clip_to_words to align bounds to words; auto_reframe for face reframing; detect_silences to cut gaps; set_default_subtitles to apply the account's caption style.
- Export. export_clip for one clip or export_all_clips for the batch, wait_export until done, get_export_frame to check a frame before downloading.
- Control. get_quota shows what is left on the month's plan.
In conversation, all of this fits in one exchange: you give the link and your criteria, the assistant chains the calls and hands back download links. Expect a few minutes of processing for an hour of video.
A prompt that works
"Import https://youtube.com/watch?v=… . When the transcript is ready, list the moments with their scores. Keep the 5 highest-scored under 50 seconds, snap bounds to words, reframe vertical, apply yellow karaoke captions at 4 words per line, export everything in 9:16 and give me the links with, for each clip, a proposed hook of 8 words maximum."
Three details that change the outcome: give a maximum length (otherwise the AI keeps clips that are too long), impose the caption style (account consistency), ask for the hook in the same message (the assistant has the transcript in front of it, that is the best moment).
REST API or MCP: which to choose?
| REST API (script) | MCP server (assistant) | |
|---|---|---|
| Who uses it | A developer | Anyone who can write an instruction |
| When | Identical, repeated flow, integrated with another system (CMS, scheduler) | Variable flow, case-by-case decisions, human control between steps |
| Setup cost | A few hours of code | A few minutes of configuration |
| Risk | Rigid: a change of criteria needs code | The assistant may misread; ask for confirmation before export |
The two coexist: at ClipOnIt, the 25 MCP tools sit on the same REST API described in OpenAPI 3.1 (51 paths, documented at cliponit.app/docs). A script for overnight runs, an assistant for special cases.
What does not automate (yet)
- The final choice of moments. The AI scores, you decide. Accounts that publish without review end up posting out-of-context clips.
- Publishing, on some tools. Opus Clip schedules publishing from its MCP; ClipOnIt delivers MP4 files to publish yourself or through a third-party scheduler.
- Rights. No tool checks that you are allowed to clip the imported video.
- The hook. The assistant proposes a good one; the best comes from someone who knows the audience.
How much does automation save?
For a weekly one-hour episode, the manual-with-a-tool routine costs about two hours a week; with an assistant driving the flow it drops to 20 to 30 minutes, almost all of it spent reviewing the shortlist and the hooks. At two episodes a week the saving is a working day a month. Below one episode a week, the setup time is not worth recovering.
Security and data
An API key gives access to your videos; treat it like a password (environment variable, never in a shared prompt). Check where files are processed: some tools do not state a location; ClipOnIt hosts on Hetzner in Finland and deletes files at the end of the plan's history window (90 days on Pro). The 2026 tools comparison lists what each vendor states.
FAQ
Can ChatGPT edit a video by itself?
Not natively. You need to connect a video tool through MCP, a GPT Action or an API. The tool processes the video; ChatGPT orchestrates.
Do I need to code?
Not for MCP: a URL, a key, and instructions in plain language. Yes for the REST API.
Which plan do I need?
At ClipOnIt, Pro ($49/month, 60 videos) or Business. At Opus Clip, a free trial then metered billing per rendered minute. Details on the pricing page and on the other vendors' pages.
Can captions and format be automated too?
Yes: caption style, words per line and export orientation are tool parameters. The right settings are in auto captions and short video length and format.