Open source · MCP

Agents don’t describe video. They cut it.

VidMCP is a production MCP server for text-prompt mattes, behind-the-subject composition, and education plates — built for Grok, Cursor, and any MCP host.

MATTE · LIVE
01
Selected work

Complex plates. Not toy demos.

Flow fields, glowing tesseracts, depth mattes, kinetic type — click any card to play.

Also: agent pipeline · all files in /assets/demos

02
Product

What the agent actually gets

Tools over stdio — not a thin FFmpeg wrapper with marketing glitter.

Perception

Text-prompt matte

Segment and track a subject from language. Mock offline; MLX / SAM when you need fidelity.

Composition

Behind-subject VFX

Grades, plates, particles composite under the person — never painted on their face.

Education

Speech-locked scenes

Math plates and narration hooks for lessons that hold structure.

System

Agent harness

Plan → perceive → compose → critic. Quality gates instead of one-shot hope.

03
Flow

Four moves. In order.

A sequence an agent can run end-to-end.

T0

Connect

Point Grok or Cursor at the VidMCP binary.

T1

Segment

Name the subject. Get a tracked alpha matte.

T2

Layer

Build the world behind them — not on top.

T3

Render

Composite, review, iterate with the harness.

04
Install

Running in under a minute

Start on mock (no GPU). Flip to MLX on Apple Silicon for real mattes.

shell
# clone + run
git clone https://github.com/dhirajlochib/VidMcp.git
cd VidMcp && python3 -m venv .venv
source .venv/bin/activate && pip install -e .
export VIDMCP_SAM_BACKEND=mock
vidmcp
MCP Server Config
{
  "mcpServers": {
    "vidmcp": {
      "command": "/path/to/VidMcp/.venv/bin/vidmcp",
      "env": {
        "VIDMCP_SAM_BACKEND": "mock"
      }
    }
  }
}
05
About
DL

Dhiraj Lochib

Full-Stack · Blockchain · AI · VidMCP

Building where agents stop describing media and start producing it. Also: fintech, smart contracts, and AI products — dhirajlochib.com

Contact

Opens mail to dhirajch145@gmail.com

MIT

Clone it. Wire Grok. Cut something.

Mock backend for first run. Real SAM when you’re ready.