中文 | English

用 Python 创建和编辑 CapCut / 剪映草稿。这是 capcut-cli 的一层薄封装：每次调用启动一次 capcut 命令，不经过 shell，返回它打印的那一份 JSON。没有服务、没有守护进程，磁盘上的草稿就是全部状态。打开剪映时，每一轨都还是可编辑的。

npm install -g capcut-cli   # 命令行本体，需要 Node ≥ 18
pip install capcut          # 本包，纯 Python，无依赖
capcut doctor               # 检查环境

import capcut

d = capcut.run("quickstart", "旁白短视频", video="clip.mp4", ratio="9:16")
capcut.run("add-text", d["draft_path"], "0s", "3s", "你好，世界", font_size=16)
print(capcut.run("lint", d["draft_path"])["summary"])

• 关键字参数就是命令行选项：font_size=16 → --font-size 16，karaoke=True → --karaoke，列表会重复该选项，None / False 直接省略。
• 位置参数原样传递，每个参数就是一个 argv，中文、空格、引号都不需要转义。
• 全部命令、参数和选项见命令参考（中文），或者在 Python 里 capcut.describe()。使用 capcut.describe(compact=True) 获取精简索引，capcut.describe(command="compile") 获取单个命令的完整契约（命令行需要 v0.28.0 或更新版本）。

命令非零退出会抛出 capcut.CommandError，带 status、data（CLI 打印的 JSON，通常含 error）、stdout、stderr：

try:
    capcut.run("lint", path)
except capcut.CommandError as e:
    print(e.status, e.data)          # lint 有错误时退出码为 2

不想抛异常就用 capcut.run_raw(...)，它返回 Result（ok、status、data、error）。找不到 capcut 命令时抛 capcut.CliNotFound，提示里有安装命令；也可以用环境变量 CAPCUT_CLI 指定，例如 CAPCUT_CLI="node /path/to/capcut-cli/dist/index.js"。

Windows 上的 CAPCUT_CLI 使用双引号包住带空格的路径（不是 POSIX 单引号）。例如在 PowerShell 中：

$env:CAPCUT_CLI = '"C:\Program Files\nodejs\node.exe" "C:\CapCut Tools\dist\index.js"'

capcut serve 是一个无状态的 JSONL 任务队列。从 Python 喂任务进去，拿回每个任务一条结果：

results = capcut.serve([
    capcut.Job("add-text", project=path, args=["8s", "2s", "关注我"], id="title"),
    capcut.Job("lint", project=path),
], workers=2)
for r in results:
    print(r["id"], r["ok"], r["status"], r["stdout"])

失败的任务是一条 ok: false 的结果，不是异常。

新建的草稿是明文，据报告剪映 11.4（macOS）能打开并就地升级，其他版本未验证；已有的加密草稿本 CLI 不读取。capcut.doctor() 会报告环境，capcut.run("decrypt", path) 会报告某个草稿的加密状态；来龙去脉见 jianying-encryption.zh-CN.md。

• 在用 Python 驱动 CapCut / 剪映？到 这个讨论 说说你在做什么，这决定 Python 客户端下一步做什么。
• 与赞助无关：如果你正在把 capcut 集成进自己的产品，或者需要它实现目前还不支持的功能，我会承接少量集成项目。请发邮件至 [email protected]，写明你在做什么。
• 想支持这个项目：成为赞助者。

Create and edit CapCut / JianYing drafts from Python. A thin layer over capcut-cli: each call spawns the capcut binary once, without a shell, and returns the one JSON document it prints. No server, no daemon; the draft on disk is the only state, and every track stays editable in the app.

npm install -g capcut-cli   # the CLI itself, Node >= 18
pip install capcut          # this package, pure Python, no dependencies
capcut doctor               # environment check

import capcut

d = capcut.run("quickstart", "Narrated short", video="clip.mp4", ratio="9:16")
capcut.run("add-text", d["draft_path"], "0s", "3s", "Hello, world", font_size=16)
print(capcut.run("lint", d["draft_path"])["summary"])

• Keyword arguments are flags: font_size=16 → --font-size 16, karaoke=True → --karaoke, a list repeats the flag, None / False are dropped.
• Positional arguments pass through as they are, one argv token each: text with spaces or quotes never needs escaping.
• Every command, argument and option: command reference, or capcut.describe() from Python. Use capcut.describe(compact=True) for the small discovery index and capcut.describe(command="compile") for a complete command contract (requires CLI v0.28.0 or newer). A list selects several names.

A non-zero exit raises capcut.CommandError with status, data (the CLI's JSON, usually with error), stdout, stderr:

try:
    capcut.run("lint", path)
except capcut.CommandError as e:
    print(e.status, e.data)          # lint exits 2 on errors

capcut.run_raw(...) never raises; it returns a Result (ok, status, data, error). A missing binary raises capcut.CliNotFound with the install line; CAPCUT_CLI can point at one explicitly, e.g. CAPCUT_CLI="node /path/to/capcut-cli/dist/index.js".

On Windows, quote paths with double quotes inside CAPCUT_CLI. For example in PowerShell:

$env:CAPCUT_CLI = '"C:\Program Files\nodejs\node.exe" "C:\CapCut Tools\dist\index.js"'

When setting the variable from Python, subprocess.list2cmdline([node_path, cli_path]) builds Windows quoting; shlex.join(...) is for POSIX. Command arguments are passed directly to the child process without shell expansion.

capcut serve is a stateless JSONL job queue. Feed it jobs from Python and get one result per job:

results = capcut.serve([
    capcut.Job("add-text", project=path, args=["8s", "2s", "Subscribe"], id="title"),
    capcut.Job("lint", project=path),
], workers=2)
for r in results:
    print(r["id"], r["ok"], r["status"], r["stdout"])

A failed job is a result with ok: false, not an exception.

• Driving CapCut or JianYing from Python? Tell us what you are building in this discussion; it decides what the Python client gets next.
• Separate from sponsorship: if you are building capcut into a product, or you need it to do something it does not do yet, I take on a small number of integration engagements. Write to [email protected] and say what you are building.
• To support the project: become a sponsor.

cd python && python -m unittest discover -s tests -v
python -m build

MIT, same as capcut-cli.