CapCut MCP Server 🎬✨

A Model Context Protocol (MCP) server for CapCut Desktop. Enables AI coding agents (Claude Code, Cursor, Windsurf, Antigravity) to programmatically assemble vertical videos, remove backgrounds via CapCut's native AI cutout, auto-duck background music, punch-zoom for retention, style viral karaoke captions, and lint social media safe zones directly on your local timeline.

JSON in, real CapCut draft out — open CapCut Desktop and your project is 100% assembled, layered, and ready for 1-click human review and export.

Key Capabilities

1. 🪄 AI Background Removal (Smart Portrait Cutout)

• Native CapCut Cutout (flag: 3): Programmatically sets the smart portrait matting flag on any video material in draft_content.json.

• Temporal Consistency: CapCut Desktop's local GRU video matting model (tt_matting_video_gru) automatically processes the subject when opened or exported — no flickering, transparent ghosting, or manual rotoscoping required.

• Chroma Key: Green/blue screen cutout with custom hex color, intensity, and shadow parameters.

2. 🎙️ Studio Audio Improver & Auto-Ducking

• Fast DSP Mode (~1s):

  • High-pass filter (80 Hz) strips table thumps, traffic rumble, and HVAC infrasound.

  • FFT spectral de-hiss (afftdn) removes steady microphone hiss and fan noise.

  • Spoken clarity EQ (+2.5dB presence boost at 3.5kHz, -1.5dB muddiness cut at 250Hz).

  • Dynamic compression for level vocal consistency.

  • -14 LUFS EBU R128 Normalization: Broadcast/streaming standard for TikTok Shop, Instagram Reels, and YouTube Shorts.

• Neural Isolation Mode (Demucs):

  • Uses deep neural stem separation (htdemucs) to isolate clean vocal stems from room echo, background chatter, and music bleed.

• Smart Music Auto-Ducking:

  • Injects KFTypeVolume keyframes to drop background music to 12% during speech and swell it to 35% during pauses and transitions.

3. 🔍 Retention Punch Zooms & Motion

• Dynamic Attention Resets: Automatically alternates between wide (1.0x) and punch-in (1.15x) keyframes (UNIFORM_SCALE) every 3–4 seconds to prevent viewer drop-off.

4. 🟡 Viral Karaoke Word-by-Word Highlight Captions

• Hormozi / MrBeast Typography: Takes word-level timestamps (Whisper, ElevenLabs) and highlights the active spoken word in #FFD700 (yellow) or #00FF66 (green) with a heavy black outline (#000000FF).

5. 🖼️ PiP Proof Overlays & Hook Headline Badges

• Proof Overlays: Injects Amazon 5-star reviews, scientific study screenshots, or TikTok comment bubbles into corners (top_right, top_left, center_pop) on Track 2 with smooth scaling and drop shadows.

• Hook Badges: High-contrast headline banners (e.g. "3 WARNING SIGNS", "DOCTOR REVEALS") pinned to top-center.

6. 🛡️ Social Media Safe Zone Linter & Fast Proxy Previews

• Safe Zone QC (capcut_lint_safe_zones): Validates that text and overlays do not collide with TikTok / Instagram Reels UI elements (right-side like/comment icons, bottom username/sound banner).

• Fast Proxy Renderer (capcut_render_preview_proxy): Stitches a lightweight 540x960 proxy MP4 in 2–4 seconds via FFmpeg for instant headless review.

7. 🌐 Multi-Language Cloner & Dubber

• Instant EN $\to$ ES Localization: Clones an existing draft, keeps all visual cuts, B-roll timing, and cutout settings, while swapping the audio track and regenerating translated captions.

Related MCP server: CapCut MCP

Tools Reference (17 Tools)

Setup Instructions

Prerequisites

• Python 3.10+

• FFmpeg & FFprobe on system PATH

• CapCut Desktop (Windows or macOS)

Installation

git clone https://github.com/bchenner/capcut-mcp.git
cd capcut-mcp

# Install base dependencies
pip install -e .

# (Optional) Install neural voice isolation dependencies
pip install -e ".[neural]"

Connecting to AI Agents

1. Claude Code

Add to your project's .mcp.json:

{
  "mcpServers": {
    "capcut": {
      "command": "python",
      "args": ["-m", "capcut_mcp.server"],
      "env": {
        "PYTHONPATH": "/path/to/capcut-mcp/src"
      }
    }
  }
}

Or add globally via Claude Code CLI:

claude mcp add capcut -- python -m capcut_mcp.server

2. Cursor / Windsurf

Add to your mcp.json settings:

{
  "mcpServers": {
    "capcut": {
      "command": "python",
      "args": ["-m", "capcut_mcp.server"],
      "env": {
        "PYTHONPATH": "C:/path/to/capcut-mcp/src"
      }
    }
  }
}

Python API Example

from capcut_mcp import CapCutDraft, improve_audio, render_draft_preview

# 1. Enhance speech audio
res = improve_audio("voiceover.wav", mode="fast", target_lufs=-14.0)

# 2. Build 9:16 vertical project
draft = CapCutDraft("Viral_Supplement_Ad", width=1080, height=1920, fps=30)

# 3. Add talking-head avatar with AI background removal
clip = draft.add_video_clip("avatar.mp4", start_ms=0, duration_ms=15000, smart_cutout=True, track_index=1)

# 4. Add dynamic punch zooms every 3.5s
draft.add_punch_zooms(clip["segment_id"], zoom_scale=1.15, interval_s=3.5)

# 5. Add B-roll background
draft.add_video_clip("broll.mp4", start_ms=0, duration_ms=15000, track_index=0)

# 6. Add background music with auto-ducking
music = draft.add_audio_clip("music.mp3", start_ms=0, duration_ms=15000, track_index=1)
draft.add_audio_ducking(
    music_segment_id=music["segment_id"],
    speech_intervals_ms=[{"start_ms": 1000, "end_ms": 6000}, {"start_ms": 7000, "end_ms": 14000}],
    duck_volume=0.12,
    normal_volume=0.35
)

# 7. Add PiP Amazon 5-star review proof overlay
draft.add_proof_overlay("amazon_5star.png", start_ms=3000, duration_ms=4000, position="top_right")

# 8. Add top hook warning badge
draft.add_hook_badge("WARNING: HIGH CORTISOL", start_ms=0, duration_ms=3500)

# 9. Lint for TikTok UI safe zones
report = draft.lint_draft(platform="tiktok")
print("Safe zone check:", report["passed"])

# 10. Save to CapCut Desktop
draft.save()
print("Draft ready in CapCut Desktop!")

Testing

Run the comprehensive test suite verifying FFmpeg audio mastering, draft generation, motion keyframes, auto-ducking, karaoke captions, safe-zone linter, and proxy rendering:

python tests/test_mcp.py

License

MIT License © 2026 bchenner