给 agent 的
一个会剪片的 MCP server。没有你,什么都出不去。
把你的 agent 指向 vidcut,它拿到的是一条真的时间轴,不是一个对话框。它同时也拿到你——在浏览器里,渲染之前先放行。
更新于 2026-09-10
什么是 MCP 视频编辑器?
Model Context Protocol(MCP)是把 AI 模型接上外部工具的开放标准,Claude Code、Claude Desktop、Cursor 等 agent 客户端都在用。vidcut 内置一个 MCP server:它不是在编辑器旁边塞一个对话窗口,而是让你的 agent 直接对一条真的、可编辑的时间轴(竖屏、横屏、方形、4:5)调用 53 个结构化工具——导入素材、剪裁与重排 clip、上逐字字幕、加动态图形、混音、用 FFmpeg 渲染。每个工具都返回结构化的项目状态,agent 不必靠逐字稿重新描述视频。而且每一次改动都落在带人工审核闸的浏览器界面上:agent 可以调用 request_review 停下来等,你放行、或带着批注退回让它接着改——没有你签核,什么都不会渲染。自部署版依 AGPL 免费;云端版多了渲染、存储与更多工具。
不是每个视频 MCP server 都这样运作。
vidcut 以 AGPL 开源,跑在你自己的机器上。它开放出来的是一条真的、可编辑的时间轴,不是一个渲染队列;而且它有一道 review gate——agent 调用 request_review 之后会停下来等,你可以带着批注把工作退回去,它读得到。你不是在放行一个已经完成的文件,你在剪的当下就在里面。
怎么接
云端
一行命令。每个付费方案都含一个 Remote MCP 端点。
claude mcp add --transport http vidcut https://your-editor.studio.usevidcut.com/mcp53 个工具
每个工具都返回结构化的项目状态,所以你的 agent 不必再把视频内容重讲一遍给自己听。
| 工具 | 做什么 |
|---|---|
add_clip | Append already-imported media to the end of the main track (existing clips are untouched — good for adding one file at a time) |
add_mograph云端 | Add one parameterized motion-graphics overlay (the others are untouched) |
add_overlay | Add a single overlay (the others are untouched) |
add_to_library | Save media into the user's cross-project library for reuse in future projects |
auto_caption | One-shot auto captions: recognize → paginate → write the caption track (a whole-track replace |
export_publish_package | Package the finished render for manual upload: copies the output video plus cover and .srt into output/publish/<stamp>/, writes one text file per platform from the metadata you provide, and records per-platform duration/size warnings in manifest.json |
extract_audio | Extract a clip's audio into a standalone audio item (the clip's own volume goes to zero), so it can be adjusted, faded or deleted on its own |
generate_image云端 | Generate NEW images with AI via fal |
generate_music云端 | Generate a NEW background-music track with AI via fal |
generate_sound云端 | Generate a NEW custom sound effect with AI (ElevenLabs SFX v2 via fal) |
get_chat只读 | Read the conversation with the person reviewing your work (both sides, oldest first) |
get_editing_skill只读云端 | Fetch one editing playbook in full (markdown): workflow with tool-call sequences, silent-failure pitfalls, final checklist |
get_editor_context只读 | The user's current selection in the UI, playhead position, and dragged time range |
get_feedback只读 | Summary of human changes since the given version — how the AI reads back what the user adjusted |
get_frame只读 | Extract the frame at a given time as JPEG — the AI's eyes |
get_generation_status云端 | Check AI generation jobs |
get_history只读 | Recent change log (version/label/source/ts) |
get_mograph_snippet只读云端 | Fetch one library item's full HTML/CSS/GSAP source plus an adaptation checklist for the html mograph template (strip the page scaffold and every <script src>, register the paused window.__timelines["html-mograph"] timeline, re-layout 1920x1080 blocks for the bake canvas, swap webfonts and bundled images) |
get_project只读 | Trimmed overview of the project (clips/captions/media/version/review) |
import_from_library | Import a library asset into this project: its derivatives are copied in (no re-processing) and the file is referenced in place from the library (content-addressed, so the reference never breaks) |
import_media | Register a media file |
import_stock_media云端 | Download one search_stock_media result into the project and register it as media (builds proxy/filmstrip/peaks |
list_editing_skills只读云端 | List the built-in editing playbooks — 8 workflows covering speech editing, transcript-driven assembly, style mimicry from a reference video, beat-synced cutting, music & mixing, long-to-short repurposing, short-form packaging and pre-render checks |
list_library只读 | Search the user's cross-project asset library (reusable logos, intros, BGM…) |
list_source只读 | List importable files in a footage folder (no recursion, hidden files excluded, whitelisted extensions only) |
post_chat | Say something to the person reviewing your work — it shows up in the Chat tab of the editor and they can reply there |
preflight_check只读云端 | Read-only self-check of the current timeline, to run right before render |
redo | Redo the last N undone edits (the inverse of undo) |
remove_audio | Remove one audio item |
remove_clip | Remove one clip (the magnetic main track closes the gap, so every later clip shifts left) |
remove_mograph云端 | Remove one mograph overlay |
remove_overlay | Remove one overlay |
render | Render the project to a final mp4 (the project canvas, currently 1080×1920, re-encoded) |
reorder_clips | Reorder the main-track clips (order is a permutation of clipIds) |
request_review | Ask the user to review the current timeline in the browser UI |
search_mograph_library只读云端 | Search the vendored motion-graphics snippet library (373 HTML/CSS/GSAP items from the hyperframes registry), organised into 13 categories (lower-third / text-type / chart-data / map-geo / social-mock / device-ui / transition / vfx-overlay / code-dev / handwriting / logo-brand / background / camera-motion) |
search_stock_media只读云端 | Search free stock-media platforms (Pexels, Pixabay, Unsplash, Freesound) for B-roll, photos, sound effects and music beds |
set_audio | Set the whole audio track (voiceover/BGM |
set_canvas | Change the canvas size (the frame everything is composed into, and the shape of the exported video) |
set_canvas_fit | What to do when the media does not fill the canvas: contain = black bars, blur = a blurred, scaled-up fill (blur is the better look for landscape footage on a 9:16 canvas) |
set_captions | Replace the whole caption track (every existing caption is replaced |
set_cover | Set the cover image: when a rendered file is still around, the frame at that time is taken from the export (overlays and captions included) |
set_overlays | Replace the whole overlay track (every existing overlay is replaced |
set_timeline | Set the whole video main track (the initial cut) |
timeline_op | Rough-cut at a point on the timeline: split, deleteBefore (drop the picture before that time), deleteAfter (drop what follows), or freeze (insert a freeze-frame) |
transcribe | Run speech recognition (whisper.cpp) over the whole timeline and return per-word timestamps |
undo | Undo the last N edits (cursor-style: call it repeatedly to keep walking back) |
update_audio | Adjust one audio item (volume, fades, timing, ducking) |
update_caption | Change one caption only (text/start/duration/style/tokens/reveal/motion) |
update_clip | Change one clip's in/duration/volume/label/transition (its position in the main track is untouched) |
update_library_asset | Rename or retag a library asset (label/tags are what list_library searches) |
update_mograph云端 | Change one existing mograph overlay: props (shallow merge — send only the fields you want to change), duration, position, start or anchor |
update_overlay | Change one overlay only (start/anchor/duration/position/text) |
这张表由 checked-in 的工具面生成,所以不会跟 server 实际开放的内容分岔。你自己搭的开源版有其中 39 个;标成「云端」的需要付费方案。