CLI Anything Capcut — Development skill for Claude Code

cli-anything-capcut

[](https://github.com/juliang8507/cli-anything-capcut/actions/workflows/test.yml) [](LICENSE) [](https://www.python.org/downloads/) [](#requirements)

English · **[한국어](README.ko.md)**

**A CLI that lets AI agents drive CapCut video editing directly.** Agents like Claude Code and Cursor can build, validate and render CapCut / JianYing (剪映) drafts through 50+ commands with `--json` output — no GUI automation required. Built on top of `pyCapCut`, it manages an event-sourced session and offers both one-shot commands and an interactive REPL.

▶ **[Demo video — "AI controls CapCut" (2m38s, Korean narration)](https://www.youtube.com/watch?v=pgx5QqAt8Is)**

Why a CLI instead of an MCP server

When agent tooling is exposed as an MCP server, the tool schema occupies context permanently. A CLI is invoked only when needed and its output is plain text the agent can parse directly. That is why every command here takes `--json`, and why agent-facing commands such as `agent status` return *what to do next*, not just what happened.

There are external benchmarks reporting that CLI-shaped agent tooling is substantially more token-efficient than MCP (ScaleKit 2026, Jannik Reinhard 2026). Note these compare the **tool-exposure styles in general — they are not measurements of this project.**

**Version 0.5.5** · See [CHANGELOG.md](CHANGELOG.md).

Requirements

**Windows only.** This is not a portability oversight — the tool resolves fonts by scanning four Windows locations (CapCut cache, CapCut SystemFont, user Fonts, Windows Fonts), auto-detects the CapCut draft folder, and drives the CapCut GUI for rendering.