A video localization factory that speaks fluent agent

What it does

KrillinAI is a Go-based video localization workbench that ingests videos from YouTube or local files, transcribes speech via Whisper variants, translates subtitles through any OpenAI-compatible LLM, generates dubbed audio via TTS or voice cloning, and renders the final cut in both landscape and portrait layouts. It targets the full stack of platform requirements for Bilibili, TikTok, YouTube, and others, including automatic cover generation.

The interesting bit

The project treats AI agents as first-class citizens: every stage of the pipeline is exposed as a discrete CLI command that emits structured JSON and a manifest file, so agents can compose, resume, and reuse artifacts without parsing human-oriented documentation. A dedicated skills/ directory provides stable contracts for agent invocation.

Key highlights

• Modular pipeline stages (subtitle, tts, render-horizontal, render-vertical, cover) that can run solo or chain together
• Supports local Whisper variants (FasterWhisper, WhisperKit for Apple Silicon, WhisperCpp) and cloud ASR to keep transcription offline or on-premise
• LLM-agnostic translation via any OpenAI-compatible API, including local models
• Desktop, web UI, and headless CLI modes across Windows, Linux, and macOS
• Explicitly optimized for short-form vertical video and platform-specific formatting

Caveats

• The desktop release is explicitly flagged as having “some bugs that are continuously being updated,” while the non-desktop web UI is described as more complex to configure
• macOS users must manually trust the application via terminal because the desktop build is unsigned and cannot be double-clicked to run

Verdict

Content creators and automation engineers who need to repurpose video across languages and platforms will find this a credible all-in-one factory; purists who prefer best-of-breed point solutions for each stage will likely find the bundled approach overbearing.