VidMCP

Agents don’t describe video. They cut it.

MCP server for agent video editing — mattes · behind-the-subject VFX · audio polish · captions · export
Built by Dhiraj Lochib · vidmcp.com · uv tool install vidmcp

---

Creator polish pipeline

One call for a publish-ready talk-head:

orient → denoise → BGM → captions → optional BG replace → export preset

Tool	What it does
run_talking_head_polish	Full recipe in one shot
process_audio	Denoise · gate · EQ · loudnorm
mix_bgm	Ambient pad under voice (ducking)
transcribe_and_caption	Whisper words + brand burn-in
replace_background	Matte + space / blur / solid plate
smart_cut_hesitations	Dead air & filler removal
export_render	youtube_16x9 · reels_9x16 · square_1x1
import_video	Bakes portrait rotation upright by default

from vidmcp.tools.creator import run_talking_head_polish

print(run_talking_head_polish(
    "talk.mov",
    preset="reels_9x16",
    bg_mode="space",   # none | space | blur | solid
    strength=0.75,
    bgm_volume=0.35,
))

MCP agents can also call apply_recipe(..., recipe_name="talking_head_polish") or free-text run_intent(...).

Agent context (tool packs + compact results)

Agents drown when MCP hosts load ~90 tool schemas and fat JSON. VidMCP defaults to a small surface:

Setting	Default	Purpose
VIDMCP_TOOL_PACK	talking_head	Expose only the tools for that product path
VIDMCP_COMPACT	1	Drop heavy keys / truncate long payloads
VIDMCP_MAX_RESULT_CHARS	4000	Hard cap on serialized tool results

Packs: talking_head · education · vfx · admin · all

run_intent("polish talk head for reels, space bg, cut fillers", video_path="talk.mov")
project_brief(project_id)          # prefer over full get_project
list_tool_packs() / set_tool_pack("vfx")
get_project(id, detail=true)       # full manifest only when needed

Recipes: talking_head_polish · talking_head_reels · talking_head_space · talking_head_tight · talking_head_infographics.

For full surface (debug): VIDMCP_TOOL_PACK=all.

---

Gallery

_{1.1 Creator polish — orient · denoise · BGM · captions · export}

Behind the subject	Tesseract · 4D trails
Flow field	Unit circle

Sample	GIF	MP4
Creator polish	gif	mp4
Flow field	gif	mp4
Tesseract	gif	mp4
Behind the subject	gif	mp4
Kinetic	gif	mp4
Unit circle	gif	mp4
Pipeline	gif	mp4

---

Install

Needs: Python 3.11+, ffmpeg, uv recommended

uv tool install vidmcp
vidmcp --doctor

# or
pip install 'vidmcp[creator]'   # + faster-whisper + mediapipe

MCP (Grok / Claude / Cursor)

# Grok
grok mcp add vidmcp \
  -e VIDMCP_SAM_BACKEND=mock \
  -e VIDMCP_WORKSPACE_ROOT=$HOME/vidmcp-workspaces \
  -e VIDMCP_TOOL_PACK=talking_head \
  -e VIDMCP_COMPACT=1 \
  -- uvx vidmcp

{
  "mcpServers": {
    "vidmcp": {
      "command": "uvx",
      "args": ["vidmcp"],
      "env": {
        "VIDMCP_SAM_BACKEND": "mock",
        "VIDMCP_WORKSPACE_ROOT": "/path/to/workspaces",
        "VIDMCP_TOOL_PACK": "talking_head",
        "VIDMCP_COMPACT": "1"
      }
    }
  }
}

Optional real matte on Apple Silicon:

pip install 'vidmcp[mlx]'
export VIDMCP_SAM_BACKEND=mlx
export VIDMCP_MLX_MODEL_ID=mlx-community/sam3.1-bf16

---

Core product

Segment	Text-prompt subject matte (SAM / MLX / MediaPipe / mock)
Compose	Effects behind the subject
Audio	Denoise, enhance, BGM duck, loudnorm
Captions	Whisper timeline + brand burn-in
Educate	Math plates + speech-locked scenes
Export	16:9 · 9:16 · 1:1 without stretch

create_project → import_video (orient)
  → process_audio → mix_bgm
  → segment_subject | replace_background
  → transcribe_and_caption → export_render

Prefer high-level paths: run_talking_head_polish · apply_recipe · run_quality_gated_pipeline.

---

Project layout

src/vidmcp/     MCP server, creator tools, perception, effects
demos/samples/  GIFs · MP4s for README & site
site/           vidmcp.com (static Hostinger)
configs/        Claude / Cursor snippets

---

Author

Dhiraj Lochib — Full-Stack, Blockchain, AI
dhirajlochib.com · GitHub · LinkedIn · [email protected]

---

License

MIT