cocut
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@cocut add a lower third at 0:10 for Ada Lovelace, role Analyst"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Cocut
An open-source video editor where you and an AI agent edit the same timeline.
Cut, trim, layer, caption and mix in the browser while Claude Code, Codex or any MCP client edits the same project. Every change, yours or the agent's, appears live, is labelled with who made it, and can be undone. Motion graphics are generated by a built-in engine and can be split into separately editable layers.

Why
Agents are good at precise, repetitive, describable edits: "make every caption 52px", "shift everything after the chart by half a second", "add a lower third at 0:10". People are good at taste: what looks right, what feels late. Cocut puts both on one timeline:
One source of truth. The edit is a plain
project.json. The editor, thecocutCLI and the MCP server all read and write it, atomically and versioned.Live. When the agent changes something, the timeline updates in your browser within a quarter-second. Agent clips are marked orange.
Undoable. Every change from anyone is a snapshot.
⌘Zundoes the agent's last step as easily as your own.
Related MCP server: artdaddy
Features
Editing
Video, image, generated, audio and text tracks; any number of each. The first track is drawn on top.
Drag media from the bin, or drop files anywhere to import. Move clips, including across tracks, and trim their edges. Split at the playhead, delete, ripple delete, copy and paste. Snapping to clip edges, the playhead and markers.
Marquee select: drag across empty space.
⌘A, shift- or cmd-click, ⌥-click selects a clip and everything after it on its track.Batch edit: with many clips selected, the inspector edits them all, showing mixed where values differ. "Make all look like the first", group and ungroup (
⌘G), nudge, and find & replace across text.Per-clip fades, opacity, position, scale and speed. Audio gain, fades, ducking under a voice track, and an optional musical ending, where the track lands on a downbeat and ends on a hit.
Text clips with size, width, alignment, colour and word-by-word highlight. Drag text directly on the picture.
Thumbnails and waveforms on the timeline, markers, zoom, and full keyboard control (
?lists the shortcuts).English and 中文 interface.
Generated motion graphics
Built-in templates:
title,numbers(count-up with a grid of squares),lowerthirdandbars. Each is a small JavaScript file incocut/engine/templates/, so adding one is easy.Split into editable elements. A generated clip can be re-rendered as separate layers: background, graphics, counters and text, each with transparency (VP9 + alpha) on its own track. Its text can come back as native text clips you can retype, restyle and retime.
Agents
In the editor: the AI tab sends a request to Claude Code (headless
claude -p) or Codex (codex exec). Its commands and replies stream into the chat as it works.From a terminal: any agent can use the
cocutCLI; see AGENTS.md.Over MCP:
cocut mcpexposes the editor as tools to Claude Code, Codex and other MCP clients.
Export
H.264 + AAC MP4 at −14 LUFS, rendered by ffmpeg with exactly the same layout rules as the preview.
Install
Requires Python 3.9+, ffmpeg on your PATH (with libvpx-vp9 for transparent layers) and Chromium for Playwright (motion graphics).
git clone https://github.com/csy2443470971-ui/cocut && cd cocut
pip install -e . # or: pip install numpy scipy soundfile pillow playwright
python3 -m playwright install chromium
pip install librosa # optional: musical endings for music clipsQuick start
cocut new my-video --size 1920x1080 # or: python3 -m cocut.demo (builds a demo project)
cocut serve # → http://localhost:8770Projects live in ./cocut-projects/ by default (set COCUT_WORKSPACE to change it).
Working with agents
Claude Code, inside the editor: open the AI tab, type what you want, and press ⌘↩. Requires the claude CLI.
Claude Code over MCP:
claude mcp add cocut -- python3 -m cocut.cli mcpCodex over MCP (~/.codex/config.toml):
[mcp_servers.cocut]
command = "python3"
args = ["-m", "cocut.cli", "mcp"]Any agent with a shell:
cocut show # timeline with clip ids
cocut set size=52 y-=0.03 --kind text --by claude # batch edit every text clip
cocut gen lowerthird V2 10 --params '{"name": "Ada Lovelace", "role": "Analyst"}'
cocut layers <clip> --text-clips # split a generated clip into editable elements
cocut shift 0.5 --range 7: --by codex # move everything after 7 s
cocut undococut -h lists every command and the selector flags (--ids, --track, --kind, --range, --text, --group, --author).
How it works
browser editor ──┐ ┌── cocut CLI (people, Claude Code, Codex …)
├──► project.json ◄───┤
MCP clients ─────┘ (atomic, versioned) └── server: watches the file → pushes changes to every open editor
│
export.py → ffmpeg → MP4 engine/ → headless Chromium → generated clips / layersMore detail in docs/architecture.md.
Project format
{
"meta": { "name": "demo", "width": 1920, "height": 1080, "fps": 30, "version": 12 },
"tracks": [ { "id": "T1", "kind": "text", "name": "Text" }, { "id": "V1", "kind": "video", "name": "Video" },
{ "id": "A1", "kind": "audio", "name": "Voice", "role": "voice" } ],
"clips": [ { "id": "c1a2b3", "track": "V1", "kind": "mg", "src": "assets/generated/g1/title.mp4",
"start": 0, "in": 0, "out": 3, "opacity": 1, "fadeIn": 0, "fadeOut": 0, "x": 0, "y": 0, "scale": 1, "by": "human" },
{ "id": "c4d5e6", "track": "T1", "kind": "text", "text": "Hello", "start": 0.3, "dur": 2.4,
"x": 0.06, "y": 0.86, "w": 0.88, "size": 58, "color": "#ECE6EA", "karaoke": true, "words": [ { "w": "Hello", "t": 0.5 } ], "by": "claude" } ],
"markers": [ { "t": 3, "label": "Numbers" } ],
"log": [ { "at": 1760000000.0, "by": "claude", "what": "set size on 3 clip(s)" } ]
}Status
Version 0.1. The editing model, the CLI, MCP, layer splitting and export are tested end to end. Still to come:
transitions beyond crossfades
keyframed properties
audio fades drawn on clips
multi-user (several people at once)
previewing a music ending before you export
Contributions are welcome; see CONTRIBUTING.md.
License
MIT. Any music, footage or fonts you put into a project keep their own licences.
Available Tools
4 toolsassetsC
Media in the project that can be placed.
Inputs: no named parameters.
helpC
Every cocut command and the selector flags.
Inputs: no named parameters.
runB
Run any cocut command (without the leading 'cocut'), e.g. "text 'Hello' T1 2 3 --y .8", "set size+=4 --track T1", "split 12.5", "gen title V2 0 --params '{"text":"Hi"}'", "layers c1a2b3 --text-clips", "undo". See help.
Inputs: command (required).
showA
The timeline: tracks, clip ids and times. Call this first.
Inputs: no named parameters.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 4 tools
Each tool has a distinct primary role: help for command discovery, show for timeline state, assets for media inventory, and run for executing commands. The broad run tool could theoretically duplicate show/assets/help, but the descriptions frame those as dedicated convenience tools, leaving only minor boundary ambiguity.
All four names are single lowercase tokens, so casing and style are consistent and readable. They do not follow an action-object verb_noun pattern, mixing noun-like (help, assets) and verb-like (show, run) names, but this is a minor deviation for a CLI wrapper.
Four tools is well-scoped for a CLI-style wrapper: three small helpers plus one universal command executor. No tool feels redundant, and the count avoids bloating the surface with one tool per command.
Because run can execute any cocut command, the surface provides a universal escape hatch for timeline edits, generation, layering, splitting, and undo. help exposes command syntax, while show and assets cover timeline and media inspection, leaving no obvious dead ends.
Related MCP Connectors
AI video editor for agents and humans: timeline, captions, color, audio and generation as MCP tools.
Compose, edit and render video with your AI agent. Turn footage into finished edits with content-aware reframing, word-timed captions and motion graphics. Hosted MCP with OAuth sign-in; cloud processing uses credits.
The agentic video editor for ChatGPT and Claude. Ask in your chat and it cuts, captions, adds music, titles and motion graphics on a real timeline you can open and adjust.
AI video editor your agent can drive: plan scenes, edit a multi-track timeline, and render MP4s.
Related MCP Servers
- AlicenseNot gradedqualityCmaintenanceEnables AI agents and MCP clients to programmatically edit video projects on a local desktop editor, with 119 tools for multitrack editing, effects, captions, audio, and batch auto-editing, producing reviewable and reversible real timeline edits.AGPL 3.0
- AlicenseNot gradedqualityBmaintenanceEnables AI agents to drive a desktop video editor through MCP, providing the same editing operations as the user — placing, trimming, rippling, grading, captioning, and exporting — on the actual project with local rendering.8GPL 3.0
- FlicenseNot gradedqualityAmaintenanceEnables Claude Code, Codex, Hermes, OpenCode, and other MCP hosts to turn natural-language prompts over raw video clips into finished, beat-synced montage edits with music matching, subtitles, transitions, and picture-in-picture, while returning reviewable timeline digests and contact sheets before rendering.-
- FlicenseNot gradedqualityAmaintenanceEnables chat-driven video editing by letting an MCP host analyze raw clips and produce finished, beat-synced montages with matched music, subtitles, transitions, and picture-in-picture, all previewed as a timeline digest before rendering. It runs locally through agents such as Claude Code, Codex, Hermes, and OpenCode.-