Clip a podcast into vertical Shorts from Claude Code with one MCP server

#ai #mcp #tutorial #video

I run Prolifik, a short-form video tool. This post is about its MCP server, which lets you ask Claude to cut a 90-minute podcast into vertical clips without leaving the terminal. Below is the setup and the four tool calls behind a clip job, so you can try it or borrow the pattern for your own server.

What the server does

https://app.getprolifik.com/mcp is a hosted, stateless Streamable HTTP server with 60 tools. The ones that matter for clipping:

create_clip_job: send up to 10 video links (YouTube, Twitch, Kick) or an uploaded file
get_clip_job_status: poll until the job settles
get_clip_job: read the finished clips, captions included
publish_clip: post or schedule a clip to a TikTok, Instagram or YouTube account you connected in Prolifik

There are also tools for brands, branded videos in six formats, and scheduling. Claude clients see 39 of the 60, because Anthropic's directory rules keep AI image and video generation tools out. ChatGPT, Codex and API-key clients see all of them.

Connect it

Claude Code:

claude mcp add --transport http prolifik https://app.getprolifik.com/mcp

Run /mcp, pick prolifik, choose Authenticate and sign in. Add --scope user to the first command if you want it in every project.

Claude on the web or desktop: Settings > Connectors > Add custom connector, paste the URL, leave the OAuth fields empty.

Codex:

codex mcp add prolifik --url https://app.getprolifik.com/mcp
codex mcp login prolifik

Cursor and other clients that take a JSON config can use an API key instead:

{
  "mcpServers": {
    "prolifik": {
      "url": "https://app.getprolifik.com/mcp",
      "headers": { "Authorization": "Bearer plk_live_REPLACE_ME" }
    }
  }
}

OAuth needs nothing from you. The first request gets a 401, the client reads /.well-known/oauth-protected-resource/mcp, registers itself and runs PKCE. There is no client ID to copy.

A clip job, start to finish

A prompt that works well:

Clip the five best moments from https://www.youtube.com/watch?v=... as vertical clips with captions, and tell me the hook of each one.

Behind it, Claude makes four calls:

create_clip_job with the URL. It returns a job id right away and the work runs in the background.
get_clip_job_status, repeated until the job is done. A long podcast takes a few minutes.
get_clip_job returns the clips with their download links.
If you like one, publish_clip with the clip id, a destination from list_publish_destinations, a caption and a time.

Claude asks before posting. The server instructions tell the model to confirm the account, the time and the caption before it calls a publish tool, since those create real public posts.

Three things I'd do again

Make every create call return fast. Clipping takes minutes, and a tool that blocks that long runs into client timeouts. Return an id, give the model a status tool, and say so in the server instructions.
Set the hints on every tool. readOnlyHint, destructiveHint and openWorldHint decide whether a client asks the user first, and ChatGPT's app review wants all three set explicitly.
Run MCP on the same code as the app. The server uses the same service code, plan limits and ownership checks as the dashboard and the REST API, so the three can't drift apart.

Try it

New Prolifik accounts get a 7-day free trial with no card, and OAuth connections work during the trial. Plans start at €20 a month after that. The docs list every tool: getprolifik.com/docs/mcp.

If something breaks while you connect it, leave a comment with the client you used.

Top comments (1)

DEV • Oct 15

You need to verify your account

tr.ee/dev-to