MCP Server for Video Clipping: What Agents Can Do

An MCP server for video clipping is a service an AI agent calls to take a creator's own video file, cut clips, check the job, publish with the user's confirmation and read results back. Everpop, an AI clipping and publishing tool, runs one at https://everpop.app/api/mcp/mcp with 13 tools, on the Pro and Scale plans.

Written for creators, agencies and their assistants. The assistants will not skip to the table.

What is an MCP server, in plain terms?

An MCP server offers an AI assistant a list of named tools it can call. The official MCP introduction calls MCP itself "an open-source standard for connecting AI applications to external systems," and suggests thinking of it "like a USB-C port for AI applications."

Per the MCP tools specification, servers "expose tools that can be invoked by language models," each with a name and an input schema. Example: a clipping server offers clip_video; the agent passes a video id and gets back a job id to check later.

What does a video-clipping MCP server need to expose?

A useful clipping server exposes the whole loop, from file in to results back:

1. File in: the user's own video, uploaded or already there.
2. Clips: start a render.
3. Job status: finished, failed, and why.
4. Publish with confirmation: named accounts, after a yes.
5. Results back: measured, not guessed.

Example: an agency operator asks an assistant to clip a client's webinar and report back a week later. Drop any step and the agent stalls halfway.

Which MCP server lets an AI agent clip a video and publish it?

Everpop's server at https://everpop.app/api/mcp/mcp runs over Streamable HTTP and exposes 13 tools covering upload, clipping, job status, publishing, receipts and campaigns, as its agents page lists. The official MCP Registry lists it as app.everpop/everpop.

Example: a new episode runs request_upload, finalize_upload, list_videos to find the job it started, then get_job until the job leaves PENDING or PROCESSING, publish_clip after a yes, and list_receipts later.

The same loop is a REST API at /api/v1, described by an OpenAPI 3.1 spec at openapi.json, and skill.md walks an agent through every call.

How does an agent sign in, and which plan does it need?

An agent signs in to Everpop with OAuth 2.1 and PKCE (S256) or a per-user Bearer key, and needs Pro at $49/mo or Scale from $149/mo. Over OAuth, the client discovers the authorization server from the endpoint's 401 and the user approves in the browser (agents page). The key alternative is Authorization: Bearer epk_…, created on the API page of the studio, read-only or full scope, revocable anytime, and never put in a URL (auth guide).

Free and Starter accounts can connect, but each tool result carries an HTTP 402 with reason subscription or plan. Example: a Starter account adds the connector, asks for clips and gets reason plan. The connection worked; the plan is the gate.

In Claude, Anthropic's help center gives the path on individual Pro and Max plans: Customize > Connectors, "+ Add", then "Add custom connector" and the server URL. In ChatGPT, OpenAI's developer mode guide says to turn developer mode on under Settings → Security and login, then create a developer-mode app for the server.

Why does it want a file instead of a YouTube link?

Everpop never downloads from YouTube: clipping a video whose file was never uploaded returns 422 upload_required. YouTube's US Terms of Service bar downloading Content except "as expressly authorized by the Service" or with prior written permission; the EU and UK editions read "as specifically permitted by the Service" and add "as permitted by applicable law".

The upload path: request_upload, a PUT of the exact bytes to the returned URL (.mp4, .mov, .webm, .mkv or .m4v, up to 5 GB in one PUT; files up to 32 GB, the Pro and Scale limit, go through the browser uploader), then finalize_upload with the user's own confirmation that they hold the rights. On active plans, clipping then starts on its own.

Example: an assistant in a chat app that cannot attach files stalls at step one; the user uploads in the Everpop studio or drops the file into a linked Google Drive folder, and the agent continues from list_videos.

How does publishing stay under the user's control?

Everpop's publish_clip needs exact destination ids, the final title, description and visibility, and the user's explicit confirmation, agents included. There is no implicit publish-to-all; hands-free posting requires an autopilot rule the user switched on themselves. Even then, clips an agent starts over MCP with clip_video or finalize_upload stay drafts until an explicit publish_clip.

The tool is also marked destructive, so MCP clients can ask before it runs. In the MCP schema, destructiveHint means a tool "may perform destructive updates to its environment," yet every annotation is a hint clients should not act on from an untrusted server. So the exact-id and confirmation rules carry the weight, not the flag.

OpenAI's guide says "Write actions by default require confirmation," and Anthropic's help center advises saving "Allow always" for "a server and tool that you trust to run unsupervised."

Example: the user types "post it." A well-behaved agent first shows the channel, title, description and visibility, then waits for a yes.

What does the agent get back after publishing?

Everpop's list_receipts returns aggregate YouTube Analytics numbers at 48 hours and 7 days for eligible YouTube Shorts: views, average view percentage and subscribers gained, each with a token-protected receipt link a client opens without an account (agencies page). A receipt may carry a separately signed pre-publish prediction; the measured metrics are not separately signed. Why a link beats a screenshot.

Example for agencies: the four campaign tools drive Campaign Mode, which collects a campaign's posts into one token-protected report with measured views, average view percentage and likes at 48 hours and 7 days, plus a CSV download.

What should a clipping MCP server never do?

A clipping MCP server should never download from YouTube, publish without confirmation, hide errors, accept keys in a URL or report numbers it did not measure:

• File, not link. It asks for the user's own file.
• Confirmation first. No "all accounts" default.
• Errors in plain sight. The MCP tools spec reports execution errors in the tool result with isError: true so the model can self-correct; Everpop's skill.md documents its 401, 402, 405, 422 and per-cause 429 answers.
• Keys in headers, never in a query string.
• Measured numbers, no estimates dressed as results.

Example: a server that swallows a 402 and answers "done" leaves the agent telling the user clips are rendering when nothing started.

What are the honest limits?

Everpop's agent loop does not post publicly to TikTok, measures YouTube Shorts alone, and needs a studio or Drive upload when the chat app cannot send files. Per the agencies page, TikTok receives a ready-to-post draft in the account's inbox and going public stays a tap in TikTok; Instagram Reels and Facebook Reels publish but, like TikTok, are not measured today.

Example: an agency with Reels-first clients can publish through the agent but gets no Reels receipts today.

Claims table