yt-transcript is a command-line tool for retrieving English YouTube transcripts with chapter and timestamp context, along with an MCP server for compatible AI clients. It can return raw captions, previews, or Markdown edited through a configured model, and it does not download audio or video or require FFmpeg. Raw CLI output, preview operations, plain-text MCP responses, and retrieval tools make no model calls. Optional Markdown editing uses Pi by default, or direct OpenRouter and DeepSeek backends with an explicitly selected model; these requests send transcript text to the provider, can incur charges, and do not automatically retry or return partial output. The CLI supports standard, focused, and summarized editorial modes, while MCP can filter SponsorBlock segments or retrieve the full transcript. MCP also exposes video-info, segment, search, and pagination tools backed by a private local source-caption cache that is enabled by default and expires snapshots after 24 hours. Provider settings include output-token limits and optional thinking controls, but higher limits do not guarantee completion and reasoning can increase cost and latency. The pinned installation uses release `v1.1.0`, Python `3.14.7`, Deno `2.3.0+`, and a separately installed Pi runtime when the default backend is used. The documentation warns users to install from the pinned GitHub release rather than an unrelated PyPI project and to keep API keys in the process environment rather than TOML configuration.
