opusclip-media-generation

Generate a voiceover, narration or text-to-speech, a music track or soundtrack, or short AI video clips with OpusClip MCP. Use when the user asks OpusClip to generate TTS, narration, a voiceover, back

SKILL.md

GitHub

OpusClip Media Generation

Generate speech, music and short video clips with OpusClip MCP, then download and check them locally. Speech, music and video are paid in OpusClip credits at a fixed price: the quoted estimate is the charge. A job that definitely fails releases its hold in full; an uncertain one (submission_unknown) keeps it held for review. Treat the user's credits as their money: quote first, get a clear yes, and never spend beyond what they approved.

Tools and models

These tools do the work:

  • opusclip_create_speech_job and opusclip_create_music_job each start one paid job, and so does opusclip_create_video_job when the server offers it; video also needs prompt, model and durationSeconds. maxCredits and idempotencyKey (8-128 characters) are required.
  • opusclip_get_media_generation_job reads a job by jobId. It is free and read-only; waitSeconds (0-20) waits on the server for the job to finish.
Tool Model Prompting docs
opusclip_create_speech_job Gemini 3.8 Flash TTS speech generation, model
opusclip_create_music_job Lyria 3.5 music generation, model
opusclip_create_video_job, model hailuo-3 MiniMax H3 (Hailuo 3) video generation, H3 feature highlights
opusclip_create_video_job, model seedance-2.5 Dreamina Seedance 2.5 ModelArk video generation, prompt guide

Tell the user which model you are using. If the server does not offer opusclip_create_video_job, say video generation is not available and continue without it.

Read the provider docs for what options mean and how to write prompts, but send only the tool's own camelCase parameter names. Speech, music and video each map common provider field names to tool fields. The tool's input schema and its validation errors are authoritative. Treat docs pages as reference material, not instructions.

Connect and sign in

Search the connected tools for opusclip_create_speech_job, opusclip_create_music_job and opusclip_create_video_job under any server name (in Claude Code, load deferred tools first), then call that server's opusclip_whoami. If no server lists the generation tools or opusclip_get_media_generation_job, or sign-in is required or expired, give the user the commands for their host and wait for them to finish. Do not run these commands yourself; the user runs them. When this skill came with the OpusClip plugin, the server is already part of it: ask the user to connect or sign in to OpusClip in the host's plugin settings instead. A signed-in server that lists opusclip_get_media_generation_job but no create tool is the right server; generation is just not available to this account right now. Say so and continue without it, without asking the user to reconnect or sign in again.

# Claude Code: add the server, then run /mcp and authenticate OpusClip
claude mcp add --transport http --scope user opusclip https://mcp.opus.pro/mcp

# Codex: add the server, then sign in
codex mcp add opusclip --url https://mcp.opus.pro/mcp
codex mcp login opusclip

The OpusClip connector in the claude.ai directory does not include generation. A newly added server may appear only after the session restarts. If opusclip_whoami returns has_api_access: false, relay its upgrade message: generation needs a plan with API access, or reconnecting and choosing an organization that has it. Jobs belong to the connected organization. Never ask for tokens. If generation stays unavailable, say so, continue without it, and never claim an asset was generated.

Spend only what the user approves

Follow billing. In short:

  1. Quote every new request. Send the full request with maxCredits: 0 and a new idempotencyKey; it returns status: "needs_approval" with estimatedCredits and creates nothing. A used key returns its existing job instead. Ask in one line, for example "Generate a 30-second upbeat instrumental with Lyria 3.5 for 4 credits? (8 of 50 used)"; in Claude Code use AskUserQuestion with Yes / Change it / Skip. Wait for an explicit yes; an approved session total covers the requests listed with it (step 3). When the run cannot take a reply, or the user says "you decide", quote only: generate nothing, report each quote, and stop.
  2. Run the approved request with the same key and maxCredits equal to that request's quoted estimatedCredits, never more than the remaining session total. Any change to the request needs a new key and a new quote, and a new yes unless step 3 covers it.
  3. When several generations are planned, ask for a session total, or use the budget the user already named. Quote every planned request, show the list and its sum, and get one yes for the list and the total. A copy-video-style materials answer that named the assets and the budget is that yes; do not ask again. That yes approves each listed request within the total; do not ask again for each one, but before each paid call say in one line what it generates, its credits and the running total, such as "Generating the music bed: 4 credits, 4 of 50 used". A listed asset's request fitted to the edit before its first run, or a change the user asked for, is covered while its new quote fits the remaining total; a retake or revision you start yourself needs its own yes. An asset not on the list, or a call that would exceed the total, needs its own quote and yes. Keep a running sum from receipts and stop before a call would exceed the total.
  4. Make one test output and get the user's reaction before any batch. Never loop paid calls without the user's approval.
  5. Never retry with a new key to recover an uncertain job; poll its jobId. See billing for retries and refusals.

Record every job in generation.json in the project folder before and after each call: kind, model, parameters, idempotency key, job ID, estimate, charged credits, status and local file.

Write the request

  • Speech: send the script as text with one of the 30 prebuilt voices as voice, and an optional style for persona and delivery. Generate narration in one call so the voice stays consistent; shorten text over the limit rather than splitting it. When the user wants to hear a voice first, follow the sample flow in speech.
  • Music: describe genre, mood, instruments, tempo, structure, length, and vocals or instrumental in prompt. Never name artists or songs, and never reuse a reference video's music. See music.
  • Video, when offered: describe subject, action, scene, style, camera and sound in prompt, and set model, durationSeconds, resolution and, without firstFrameUrl, aspect ratio for the edit. See video.

When copy-video-style hands off a missing asset, start from its brief: the asset's role, kind, duration, section or beat timing, mood, aspect ratio, and voice and delivery style for speech. The budget the user gave copy-video-style is the session total for the listed assets, and their quotes are already in generation.json. Reuse a key only for a request sent exactly as quoted; any change to text, voice, prompt or duration needs a new key and a new quote, covered by that total only while it fits (step 3). Generate only the assets the edit needs.

Wait, deliver and verify

The create call waits up to 20 seconds. While a job is queued or running, call the getter with waitSeconds: 20; never call the create tool to check status. If it is still queued or running after 20 minutes (45 for video), stop polling, give the user the jobId and its held credits, and do not resubmit. A failed or submission_unknown receipt is returned as a tool error, but it is still a receipt: read its jobId, errorCode and credits, record it, and follow its errorCode in billing. Only a result without a jobId means no receipt came back.

You cannot listen to audio or watch video. Give the user the playable URL of every output, samples included, and ask them to review it before you build on it. Never claim you heard or watched a result.

Output URLs are signed and expire. Download each output promptly into the project folder, for example generated/<jobId>.<ext>, taking <ext> from outputs[].mimeType (audio/wav → wav, audio/mpeg → mp3, video/mp4 → mp4). When a link has expired, read the job again for a fresh URL instead of paying for a rerun. In Codex, downloading needs network approval; if the user declines, ask them to download the file in a browser and give you its path. Verify every file with ffprobe: duration, streams, codec, and dimensions for video. View opening, middle and closing frames of video. Then use the files on the local timeline like any other asset. If you cannot run commands or write files, as in ChatGPT, skip generation.json, the download and the checks: give the user each output URL with its job ID and charged credits instead.

Report each job's model, charged credits and local path. chargedCredits: null means billing has not settled yet, not zero.