You can now clip video from inside your AI agent: the state of MCP for video in 2026

Six months ago, "AI video editing" meant opening a web app, pasting a link, and clicking around. Today you can stay in the tool you already talk to (Claude, Cursor, ChatGPT) and just say: "clip the best moments from this podcast and reframe them for TikTok." The clips come back. No tab-switching, no dashboard.

The thing that made this possible is the Model Context Protocol (MCP). If you have heard the acronym but tuned out, here is the one-line version: MCP is a standard way for an AI agent to call a real tool. The agent does not guess how your API works. It reads a list of typed tools, picks one, fills in the arguments, and gets structured data back.

I have been building a video clipper (Katto) with an MCP server as a first-class surface, not an afterthought, so I went and looked at where the whole clipping category actually stands on this. Here is what I found, honestly.

Why video clipping is a near-perfect MCP use case

Most SaaS features are awkward to expose to an agent because they need a lot of back-and-forth UI. Clipping is the opposite. The whole job is: input a long video, get short clips out. That maps cleanly to a handful of tools:

• create a clip job from a URL or an upload
• poll it
• fetch the finished clips (the MP4s, the captions, the virality score)
• optionally re-render one with a different layout, or dub it into another language

That is a workflow an agent can drive end to end. You give it a YouTube link in a sentence, it hands you back five captioned vertical clips.

The landscape (as of August 2026)

MCP is becoming surprisingly common among clipping tools. OpusClip ships a hosted server with OAuth. Reap and Submagic expose MCP endpoints. Even smaller players like Whipscribe (more transcription-and-clip-search focused) run a public MCP with a local install option.

So I do not think "we have an MCP server" is much of a differentiator anymore, at least for tools targeting developers and agent workflows. What still varies a lot is how it is done. Here is the factual picture, checked against each vendor's own docs:

Tool	Hosted endpoint	Auth	Local option	MCP access tier
OpusClip	mcp.opus.pro/mcp	OAuth, no API key	hosted (mcp-remote for stdio)	free trial, then Pro (metered per-minute)
Reap	mcp.reap.video/mcp	OAuth, workspace-scoped	hosted (mcp-remote for stdio)	paid plan with API access (from ~$9.99/mo)
Submagic	api.submagic.co/mcp	Bearer key	via mcp-remote bridge	shares REST API credits
Whipscribe	whipscribe.com/mcp	Bearer key (guest tier allowed)	open-source stdio client	free tier
Katto	mcp.katto.tech/mcp	OAuth 2.1 + DCR, or bearer	npx katto-mcp	all paid plans, shared quota

Checked 2026-08-22 against vendor documentation. Things move fast in this space, if a cell is out of date, tell me and I will fix it.

What actually differs between video clipping MCPs

Access and pricing

This is where implementations diverge the most. Some tools meter agent usage: OpusClip offers a free trial and then a Pro plan priced per render-minute, and Submagic's MCP draws from the same credit pool as their metered REST API. Others include it lower down: Reap ships theirs on a paid plan with API access, and Whipscribe's works from a free tier.

On Katto, the MCP and API are included on every paid plan, and calls draw from the same monthly video quota with no per-minute surcharge. I did that on purpose, and I am not the only one who thinks agent access should not be a tax. But check the table against your own budget; "included" means different things at different price points.

Authentication

In practice, I have seen two useful authentication patterns:

• Hosted, zero-install: point an OAuth client (Claude web or desktop, or any client that supports OAuth) at the server URL and sign in. No key stored on disk.
• Local key: a one-line npx install with an API key, for Cursor, CI, and scripts.

Not everyone offers both. OpusClip and Reap are OAuth through their hosted endpoints; Submagic and Whipscribe authenticate with Bearer keys. Katto's hosted endpoint (https://mcp.katto.tech/mcp) does OAuth 2.1 with Dynamic Client Registration so the key never touches your machine, and the same endpoint also accepts a bearer key for automation. That is why I wanted Katto to support both patterns rather than forcing one authentication model on every use case.

One note on safety, because it applies to every server in the table: an MCP gives your agent real capabilities on your account. Before connecting any of them, look at what the tools can actually do, and prefer servers that keep secrets and destructive actions out of the agent's reach.

Tool coverage

A lot of MCP servers stop at "create job / get job." The more useful ones expose the whole pipeline. Katto currently exposes 15 tools, covering the full workflow from creating a job to transcripts, re-rendering with a different layout or caption style, dubbing into eight languages, and quota checks. The point is not the number, though. I wanted an agent to be able to do something useful after the first render instead of stopping at "job complete."

OpusClip goes wider still: their catalog is larger than mine and includes things Katto's MCP does not do today, like scheduling posts to connected social accounts.

What the MCP cannot tell you

Nothing in that table tells you whether the clips are any good. That is the part that actually matters, and it is the part a tool list cannot show.

Where Katto is still behind

Honesty cuts both ways, so here is the other side. OpusClip has a much more mature product around its MCP: a bigger tool catalog, social scheduling from inside the agent, and a growing pile of third-party tutorials. Katto is younger and much smaller, and some of the polish that comes from years of users banging on a product simply is not there yet. If you want the most established end-to-end agent workflow today, that is a fair reason to pick OpusClip.

The part I actually care about

Here is the honest bit. The MCP is not the product. It is a doorway to the product. What matters is what happens behind the tool call.

For Katto specifically, I have spent most of my time on videos where the interesting moment is not spoken: sports, gameplay, trailers, b-roll. Many clipping pipelines still rely heavily on transcript signals to decide what matters. That works well for podcasts and interviews, but it gets much harder when the hook is visual rather than verbal. So Katto reads the pixels too: scene cuts, faces, motion, on-screen text, instead of relying only on the transcript. That is the capability the MCP exposes; the MCP is just how an agent reaches it.

So my advice, whether you use Katto or not: do not pick a clipping tool by its MCP tool count. Pick it by what the clips actually look like on your kind of footage, then check that it has a clean, ungated, OAuth-capable MCP so your agents can drive it. Both matter. But only one is the reason someone keeps using the product.

Try it in one line

If you already have an MCP client, point it at https://mcp.katto.tech/mcp and sign in with Katto (no key on disk). Or run it locally with npx -y katto-mcp and a KATTO_API_KEY. Then just ask your agent to clip something. That is the whole point.

I am building Katto in public. If you are doing anything with agents and media, I would genuinely like to hear what tools you wish existed. Find me at katto.tech.