Skip to navigation
Plugins & ExtensionsAgent Skills

batch-lipsync skill

Lip-sync every row of a CSV or folder, resumably, within your plan’s limits or through the Batch API.

Install this skill with npx skills add synchronicity-labs/skills --skill batch-lipsync, or get every skill with the Claude Code plugin. See Agent skills.

What it does

batch-lipsync lip-syncs many videos in one go: one output per row of a CSV, a spreadsheet, a folder, or a list like “this presenter with each of these 40 audio files”. It builds a manifest, quotes the whole run, offers a pilot, then runs every row within your plan’s limits and collects the outputs into a results table.

With an API key it uses a bundled script, sync_batch.py, that uploads each file once, keeps to your plan’s concurrency, resumes after an interruption, and downloads every result. On Scale plans and above, 20 or more rows can go through the Batch API instead, either with the script or straight through the MCP server’s batch tools.

When to reach for it

What you haveReach for
Many rows of video plus audio or scriptbatch-lipsync
Personalized video per lead or customerbatch-lipsync
One recording personalized with each recipient’s name in the same voiceedit-dialogue previews, then batch-lipsync
Several videos into other languageslocalize-campaign
One videolipsync

Try

  • “Lipsync each row of leads.csv onto presenter.mp4”
  • “Make a version of this ad for each of these 12 voiceovers”
  • “Run manifest.csv through sync, the total is approved”

What you need

  • The sync. MCP server connected, or SYNC_API_KEY and Python 3.9 or later for the script. The setup skill walks you through either.
  • A list where every row has a video or image, and audio or a script with a voice.

Common questions

If a shared input is wrong, like the voice or the presenter’s framing, a pilot costs two generations instead of two hundred.

The media, the scripts, file names and the row IDs (used as output names). Columns the run doesn’t use stay local, so keep names and emails out of the row IDs.

Run the same command again. Finished rows are skipped, submitted rows are polled instead of resubmitted, and nothing already paid for is paid for again.

The script refuses to reuse that row’s result. Give the changed row a new ID, which makes it a new generation.

On Scale and above, for 20 to 500 short rows when a turnaround of up to 24 hours is fine. On other paid plans a batch can only hold as many rows as the plan runs at once, so the regular run is just as fast.

The results table lists each failure with its reason. Rows that failed on bad input need a fixed input. Retry the rest with --retry-failed.

It’s working if

  • The quote covers every row, or tells you which rows it couldn’t price.
  • No more generations run at once than your plan allows.
  • Rerunning a finished manifest creates nothing new.