Pika adds Language Swap to MCP for own-voice video dubbing

• Pika has added a Language Swap skill to Pika MCP, and pika_labs' launch post says it can swap the spoken language in a video while making it look and sound like the original speaker.
• The official Pika MCP Server page shows the tool is wired into an MCP endpoint for Claude Code, while GitHub's Pika-Plugins repo says the same MCP surface also works with Codex, Cursor, Claude Desktop, and other MCP-aware clients.
• In MayorKingAI's Japanese demo and LinusEkenstam's Mandarin demo, the output keeps the creator's voice profile and face on screen instead of falling back to a generic dub track.
• The workflow is short: MayorKingAI's setup steps and LinusEkenstam's setup steps both describe installing or updating Pika MCP, running /language-swap, uploading a talking-head clip, and choosing a target language.
• The early caveats are unusually concrete, because LinusEkenstam's disclaimer flags facial hair as a lip-sync problem and MayorKingAI's follow-up says a Japanese speaker confirmed his output sounded right.

You can add the server to Claude Code, browse Pika's MCP plugin repo, watch MayorKingAI's Japanese pass, and compare it with LinusEkenstam's Mandarin clip. The fun part is that the launch landed as a slash-command workflow, not a buried export setting, so the product pitch is basically: point an agent at your talking-head video and tell it what language you want.

Pika framed the release as a new skill inside Pika MCP, not a standalone app. In pika_labs' launch post, the company says Language Swap can change the language spoken in "any video" while keeping the speaker looking and sounding fluent.

The official Pika MCP Server page does not mention Language Swap by name, but it does confirm the broader packaging: Pika MCP is an OAuth-authenticated MCP server for AI video, image, music, and voice generation. That page also gives the exact Claude Code install command, claude mcp add pika-me --transport http https://mcp.pika.me/api/mcp.

Pika's GitHub repo adds one useful layer the landing page leaves out. It says the MCP surface exposes 58 atomic creative tools, while Pika also ships curated workflow layers as skills.

Two creator walkthroughs make the new flow pretty explicit. In MayorKingAI's setup steps, the sequence is:

1. Install or update Pika MCP in Claude Code or Codex.
2. Run the /language-swap skill.
3. Upload a video of yourself talking.
4. Keep your head inside the frame.
5. Choose the target language.

GitHub's Pika-Plugins repo says Pika's skills work across Claude Code, Cursor, Codex, OpenCode, and Cline, while the raw MCP server works with Claude Code, Cursor, Codex, Claude Desktop, and Claude.ai connectors. That matters because Language Swap is showing up as an agent-native command, not just a button inside Pika's web app.

Pika's own MCP landing page still centers Claude Code setup, and pika_labs' link post sends users straight to Pika MCP rather than to a separate dubbing tool.

The first wave of examples is all talking-head content, which makes sense for a feature that has to preserve a face, a voice, and mouth motion at the same time.

MayorKingAI says in his demo post that hearing his own voice speak Japanese felt surreal, and his prompt in the companion setup post was simply: "Make me speak Japanese in this video." A later reply in MayorKingAI's follow-up says a Japanese person confirmed the result sounded correct.

LinusEkenstam used the same workflow to dub himself into Mandarin. In his demo post, he describes the result as a quick way to tear down language barriers, while an earlier reply from LinusEkenstam points to a more automation-heavy use case for people already spending time inside Claude Code.

Min Choi's repost in minchoi's clip helped spread the launch beyond Pika's own audience, but the more useful evidence is that multiple creators got working outputs on day one with very plain prompts.

The nicest thing about these first demos is that they come with real failure notes. In LinusEkenstam's disclaimer, he says his beard caused shaking and blurring around the mouth in the dubbed version, which is exactly the sort of edge case lip-sync systems usually hide in a benchmark reel.

Both walkthroughs also repeat the framing constraint. MayorKingAI's instructions say to keep your head inside the frame, while LinusEkenstam's instructions say to keep your face inside the frame, which suggests the current sweet spot is still a clean, centered talking-head shot.

One more useful detail arrived in the replies: MayorKingAI's earlier follow-up said he was looking for a native speaker to validate the Japanese, and his later reply says he got that confirmation. That does not turn a tweet thread into a lab eval, but it is more concrete than the usual "looks good to me" launch-day reaction.