CapCut headless automation that knows exactly where its limits are

Jianying Headless is a local automation layer for the macOS desktop version of CapCut (Jianying Pro) that lets you build and modify native draft projects from structured JSON plans, without touching the application's GUI. The core idea: you describe an edit in a plan file, the tool constructs a draft that CapCut can open and continue editing natively, and only when you explicitly ask does it invoke the local rendering engine to produce an MP4.

The scope is deliberately narrow and the README is unusually honest about where the edges are. Supported capabilities include multi-track video, speed and volume adjustments, picture-in-picture, subtitles, local audio, six static geometric masks, dissolves, and a small fixed set of filters and effects that have been individually catalogued. Linear keyframe animation works; arbitrary curve-based speed changes do not. Image support covers PNG, JPEG, and GIF, with acceptance ranges documented per format combination. Compound clips are flagged as experimental, and their registration on the app's home screen is still blocked.

The version pinning is aggressive. The tool targets CapCut draft format versions 11.4.0 and 11.4.2, with native export accepting only 11.4.2. The minimum macOS version is 26.0, tested on 26.5.1. The bridge layer is compiled against Apple clang 21.0.0 with the macOS 26.5 SDK, and builds are only accepted if their output matches a fixed hash. Version numbers, bundle IDs, and full application signatures are verified before anything runs. An unsupported version stops the process entirely rather than silently degrading.

The architecture separates the engine, the bridge to native libraries, and an agent-callable Skill entry point into distinct directories. That Skill is also published separately in a public repository, but the core project remains private — installing the Skill alone gives you nothing without separately obtaining access to this repository and a matching CapCut installation. The build script compiles only the project's own bridge code and links against the already-installed application's libraries; it does not decrypt, patch, or modify them.

The licensing situation is layered and the README does not paper over it. Original code is under a personal/non-commercial license requiring written authorization for commercial use. Third-party components — including acknowledged lineage from pyJianYingDraft and jy-draftc — retain their own licenses. The repository explicitly disclaims being a clean-room rewrite or fully original work.

What makes this interesting for agent builders is the Skill interface: a JSON plan drives the whole pipeline, which means an LLM can construct or modify an edit plan and hand it off to a local tool that produces something a human can open and refine. The README is careful to note that a successfully generated file does not guarantee native playback consistency, cold-restart stability, or commercial asset clearance — those require separate verification. That kind of explicit boundary-drawing is rarer than it should be in automation tooling.