把原始素材放进文件夹里。像助理剪辑师一样和它对话。直接把.fcpxml放进时间线——每一个剪辑、过渡、重新定时和字幕的决定都附有书面理由，你可以检查、拒绝或反复修改。

又一个半吊子的AI编辑器，演示漂亮，但你一对准真实画面就崩溃了。 三条感知通道——鹦鹉ONNX语音，Florence-21帧每秒的视觉字幕，CLAP音频事件，基于代理策划的词汇评分——交织成一个merged_timeline.md，代理端到尾阅读并推理，而不是看像素。字语精确的EDL锚点来自helpers/find_quote.py，对应缓存转录——没有grep，也没有逐一切换。最先进的开放模型，以别人未曾发布的方式组成，100%在你的机器上运行。每个剪辑边界在导出前都会与源片段核实——中段剪辑和视觉断裂会阻挡FCPXML。

没有云。没有黑匣子。没有水印。没有订阅。第50版的迭代成本和第一版一样。

**这不是新鲜事。

Drop raw footage in a folder. Talk to it like an assistant editor. Get an .fcpxml straight into your timeline — every cut, transition, retime, and caption decision comes with a written rationale you can inspect, reject, or iterate on by just saying so.

Not another half-baked AI editor that demos pretty and falls apart the second you point it at real footage. Three perception lanes — Parakeet ONNX speech, Florence-2 visual captions at 1 fps, CLAP audio events scored against an agent-curated vocabulary — interleaved into a single merged_timeline.md the agent reads end-to-end and reasons over instead of watching pixels. Word-precise EDL anchors come from helpers/find_quote.py against the cached transcripts — no greps, no off-by-one. State-of-the-art open models, composed in a way nobody else has shipped, running 100% on your machine. Every cut boundary is verified against the source clips before export — mid-word cuts and visual discontinuities block the FCPXML.

No cloud. No black box. No watermark. No subscription. Iteration on revision 50 costs the same as revision 1.

This isn’t a novelty. It’s the assistant editor you wish you could hire — doing real assembly work, on your machine, in your NLE. If you edit for a living, it saves you hundreds of hours.

---

premiere-agent

Edit videos by conversation — runs 100% locally and exports straight into Premiere Pro (and Resolve, and Final Cut) with J cuts, L cuts, dissolves, and the rest of the cut vocabulary you’d actually use in an NLE.

Drop raw footage in a folder, chat with your agent (Claude Code — local CLI or cloud workspaces, or Codex CLI / IDE / app), get cut.fcpxml + cut.xml back — XML-only delivery, the cut lives in your NLE. Text + on-demand visuals, no frame-dumping — three perception lanes so the LLM can reason about speech, abstract sounds, and visible objects, not just one of them.

What it does

It’s a conversational video editor — you talk to it like an assistant editor (“trim the filler, J/L the dialogue, timelapse the build sequence, ship it”) and you get back an XML cut you import straight into your NLE.

Drop your raw clips in a folder, run claude or codex in that folder, and the agent walks the loop with you: it inventories the sources, asks a few questions about what you’re going for, runs a one-time preprocess that reduces hours of footage to three small text timelines, proposes a strategy, waits for your OK, writes the cut, self-evaluates every boundary against the source clips, and only then drops cut.fcpxml + cut.xml + master.srt next to your sources. Don’t like a cut? Tell it. Want a different pace? Tell it. The preprocess is cached, so iteration is just re-reasoning over the same timelines — no re-watching, no re-transcribing.