Open MCP · Video for agents

Don’t describe the cut. Run it.

VidMCP gives Grok, Cursor, and Claude real tools to segment, denoise, caption, replace backgrounds, and export — not a chat about FFmpeg.

Install See work GitHub
text-prompt matte behind subject denoise + BGM captions 9:16 export
REC matte · live
process_audio
mix_bgm
export_render
01
Work

Selected plates

Synthetic demos only. Click to play.

Creator polish
▶
Pipeline

Creator polish

Orient · denoise · BGM · captions · export

Behind subject
▶
Core

Behind the subject

Particles under the matte

Flow
▶
Brand

Flow field

Curl noise → logo

Tesseract
▶
Math

Tesseract

4D trails

Unit circle
▶
Edu

Unit circle

Live sin / cos

02
Product

What agents can actually do

Audio

Denoise + BGM

Clean vocals, loudnorm, ambient pad ducked under speech.

Orient

Portrait-safe

Phone rotation baked on import. Export 9:16 / 16:9 without stretch.

Matte

Behind subject

Text-prompt cutouts. Plates and particles live under the person.

Captions

Speech burn-in

Word timelines + brand styles. Infographic cards from keywords.

Edit

Smart cuts

Drop dead air and long fillers from the transcript.

MCP

Tool packs

Recipes and pack-aware paths so agents don’t wander 80 tools blind.

03
Changelog

What’s shipping

Recent product surface — no version spam, just what landed.

Latest

Agent context control

  • Default tool pack talking_head — ~20 tools instead of ~90
  • run_intent · project_brief · list_tool_packs / set_tool_pack
  • Compact results by default (VIDMCP_COMPACT=1)
  • Packs: talking_head · education · vfx · admin · all
Creator

Creator polish path

  • run_talking_head_polish — one-shot orient → denoise → BGM → captions → export
  • Flags: smart_cut · infographics · mix_bgm · process_audio · thumbnail
  • Recipes: reels · space · tight · infographics
  • Portrait rotation baked on import_video by default
Core

Matte · compose · educate

  • Text-prompt subject tracking (SAM / MLX / mock)
  • Behind-the-subject effects and plates
  • Math / education scenes and harness recipes
04
Flow

Import → export

01
import_videoBake orientation upright
02
process_audio · mix_bgmClean vocals + cinematic bed
03
replace_backgroundMatte + space / blur plate
04
transcribe_and_captionWords + brand burn-in
05
export_renderreels · youtube · square
05
Install

Ship in one line

Needs Python 3.11+ and ffmpeg. Copy any block.

zsh · install
$ uv tool install vidmcp
$ vidmcp --doctor
OK ffmpeg · server · tools
pip · extras
$ pip install 'vidmcp[creator]'
# Whisper + MediaPipe for captions & fast matte
mcp · host
$ uvx vidmcp
# VIDMCP_SAM_BACKEND=mock
# VIDMCP_WORKSPACE_ROOT=~/vidmcp-workspaces
# VIDMCP_TOOL_PACK=talking_head
# VIDMCP_COMPACT=1

One-shot polish

Tool run_talking_head_polish — full pipeline without juggling 20 calls.

Apple Silicon

pip install 'vidmcp[mlx]' then VIDMCP_SAM_BACKEND=mlx.

Source

github.com/dhirajlochib/VidMcp

06
About

Dhiraj Lochib

Full-stack, blockchain, AI. VidMCP is open MIT — agents that cut video, not just describe it.

dhirajlochib.com GitHub Email
07
Contact

Build with us

Collabs, bugs, or “I tried the MCP” — opens your mail app.