Translate one live caption stream into multiple languages at once.

August 17, 2026 · 8 min read

NotchLive already translates live captions into 20 languages with Apple Translation. That is useful, fast, and built into the app. But it is not the end of the translation story.

In NotchLive 1.9, the optional read-only local MCP connection can expose the active Timeline to an AI client such as Codex or Claude. Once connected, the model can follow each newly finalized caption and translate it into every language you ask for—in parallel, in the same continuing conversation.

Watch one caption stream become four live translations

What is actually happening?

This is not NotchLive secretly running four more translation engines. It is a clean handoff between two products that each do one part of the job well.

1. NotchLive captures the conversation. System audio, microphone speech, or both become captions through on-device WhisperKit and CoreML speech recognition.
2. The active Timeline exposes finalized text. When you separately enable active meeting captions for MCP, the connected client can retrieve each new finalized turn and its current speaker label.
3. Your AI client translates the stream. Codex or Claude checks for new revisions, translates only the new turns, and keeps producing the language outputs you requested until you tell it to stop.

NotchLive remains the caption source. The connected AI model performs the extended-language translation and displays the results in its own conversation.

Twenty built-in languages vs model-powered MCP translation

These are two different translation paths. Neither replaces the other.

So can it translate into “any language”? The practical answer is: any language the connected model can translate well enough for your use case. There is no hardcoded NotchLive MCP target list, but model coverage, quality, latency, output length, and context limits still apply. Less-resourced languages deserve extra checking.

How to set it up

1. Use NotchLive Pro and start captions normally.
2. Open NotchLive Settings and enable AI Connections.
3. Connect Codex or Claude Desktop using the setup shown on the NotchLive MCP page.
4. Separately enable active meeting captions. Saved-library access alone does not expose the live Timeline.
5. Tell the AI client exactly which languages and output format you want.

You can replace those four languages with the ones you need. You can also ask for the original text beside each translation, request a compact table, preserve product names, or tell the model not to translate specific technical terms.

The genuinely wild part is the blend

Parallel translation is already powerful. But MCP gives the AI client structured access to a changing meeting record, so translation can be combined with other instructions in the same live workflow.

• Translate and simplify: produce Japanese and German translations plus a plain-language English version.
• Translate by audience: keep a literal technical translation for engineers and a shorter explanation for customers.
• Preserve meeting structure: keep the current speaker label attached to each translated turn.
• Track important terms: preserve product names, acronyms, code words, or a supplied glossary across every output language.
• Blend translation with a live brief: translate new turns while separately tracking decisions, questions, or action items.
• Change direction mid-meeting: add another target language or change the output format without restarting NotchLive captions.

The caption stream stays singular. The connected model can turn that stream into several useful views for different people at the same time.

What NotchLive does—and what it does not expose

MCP access is off by default. AI Connections must be enabled, and active Timeline access has its own separate switch. Turning that switch off stops the live data surface.

NotchLive exposes finalized caption text, Timeline revisions, and current speaker context. It does not expose raw audio, partial captions, microphone controls, recording controls, deletion tools, or write access through MCP.

The privacy boundary changes when you involve an external AI client. NotchLive performs speech recognition locally, but Codex, Claude, or another connected client may send the requested caption text to its model provider. Use the workflow only when that is appropriate for the conversation and your policies.

Why this matters

Most live translation products ask you to choose one destination language, use a specific meeting platform, or move the conversation into a cloud workspace. NotchLive takes a more composable route: caption the audio already available to your Mac, keep the active stream under your control, and let the AI client create the language views you actually need.

That makes a small local caption app unexpectedly capable. A single call can become Spanish, Japanese, Turkish, and Arabic text at once—or a completely different combination—without adding a meeting bot and without giving the AI client control over NotchLive.