Match video lips to replacement speech

Lipsync matches the mouth movements of a real person in an existing video to replacement speech. The input can be a new voiceover, a re-recorded take, a song or a typed script rendered with a selected voice. The hosted sync. MCP workflow produces one MP4 and returns its output URL.

This single-video workflow adapts the MIT-licensed lipsync skill from synchronicity-labs/skills. It covers a single existing video. Translation, talking still photos, AI-generated characters, word-level dialogue edits and batch processing belong to separate upstream skills.

Use it to replace a talking-head clip's speech, match visible singing to a song or synchronize one human face with new words. The workflow works best with a clearly visible, well-lit face, an unobstructed mouth, at least 480p video and one speaker per track. Clean speech without music or another voice helps quality. Profiles, occlusion, low light and multiple faces can reduce the result's quality; identify these issues before spending credits. Animals are unsupported.

The workflow assumes users have rights to their media and declines requests intended to deceive viewers by making an identifiable real person appear to say words they did not say. Plan, credit or length errors stop the workflow rather than triggering repeated attempts.

A video may be a local file, public HTTPS URL, existing video asset ID or supported YouTube, TikTok, Instagram or Facebook link. Speech may be an audio file, URL, asset or a script plus voice. Ask for missing information and never invent the words. Save inputs once and reuse asset IDs.

Local media uses a signed upload URL, an upload with matching content type and asset creation that waits for completion. Social imports first create a project, then create a video asset with explicit project intent. An import_in_progress response permits an identical retry after the returned delay; import failures or limits call for an explanation and a file-upload alternative.

For scripts, list available voices and offer suitable choices unless the user already named one. If speech and video lengths differ noticeably, ask how to handle them. The sync modes are bounce, which plays forward and then in reverse by default; loop; cut_off; silence; and remap, which changes video speed. Audio trimming can address the plan's length limit. If several faces appear, ask who should speak and leave active-speaker detection unset unless someone is chosen.

Estimate cost using sync-3 unless another model was selected. Supply speech duration or the script; cut_off and silence also require video duration and sync mode for an accurate estimate. Script duration is approximate and speech synthesis is billed separately. Present the video, speech, voice, mode and estimate, then obtain approval unless that exact request and cost were already approved.

Create the generation once, retaining its idempotency and project keys before the call. A lost response or timeout permits one retry using identical keys and inputs to recover the original generation without a second charge. Check the returned job ID within a bounded wait budget until completion, failure or rejection. If still pending, report the ID and state without creating another job. Return outputMediaUrl unchanged, report rejection reasons, and ask the user to review visual quality even after completion.

The hosted server requires an account with sync, network access and a remote MCP client supporting Streamable HTTP and OAuth. For a Claude client, add the hosted server with:

claude mcp add --transport http sync https://mcp.sync.so/mcp

Sign-in happens through the client's MCP controls. Codex and Cursor can connect to the same hosted endpoint. A newly added server may require a client restart. The MCP client holds the OAuth session, so this skill does not store credentials. The REST API route is outside this entry's coverage.

Video and audio upload to sync., and generation is billed to the user's account. The free tier provides short, watermarked monthly generations; paid plans raise length, concurrency and file-size limits. New work starts a new project unless the user asks to reuse one. Another paid generation must never be used merely to check progress or refresh an output link.