Envie MCP server

Envie

AI video, verified.

Describe a video to your AI. Get a real file back.

Envie gives Claude Code and any MCP client a deterministic render engine: headless Chrome filmed frame by frame, six verification gates, and a full read-back layer. Free. No watermark. No account.

By GOL Productions.

The Problem

AI can write code. AI can describe video. But AI can't see what it made—so it guesses, you render, it's wrong, you describe what's wrong, repeat.

The Solution

You: "Make me a 15-second launch video for my app"
AI:  [writes HTML composition]
AI:  [calls envie_render]
AI:  [calls envie_see to check frames]
AI:  "Done. Video at output.mp4. All 6 gates passed."

Envie renders what your AI writes, verifies it machine-checks, and lets your AI see the result. No guessing.

Install

npx @golproductions/envie setup

Detects Claude Code, Cursor, Windsurf—registers with all of them. Then just ask:

"Make me a 15-second vertical launch video for my app"

Manual install

{ "mcpServers": { "envie": { "command": "npx", "args": ["-y", "@golproductions/envie", "mcp"] } } }

Or for Claude Code:

claude mcp add envie -- npx -y @golproductions/envie mcp

How It Works

┌─────────────┐     ┌─────────────┐     ┌─────────────┐     ┌─────────────┐
│   WRITE     │ ──► │   RENDER    │ ──► │   VERIFY    │ ──► │    SEE      │
│             │     │             │     │             │     │             │
│ AI writes   │     │ Chrome films│     │ 6 gates     │     │ AI checks   │
│ HTML comp   │     │ frame by    │     │ machine-    │     │ frames at   │
│             │     │ frame       │     │ check video │     │ timestamps  │
└─────────────┘     └─────────────┘     └─────────────┘     └─────────────┘

1. Write

Your AI reads envie_guide and writes a composition: one self-contained HTML file with CSS animations, WebGL, Canvas, GIFs—whatever the browser can render.

2. Render

envie_render films it deterministically. Headless Chrome with a virtualized clock: performance.now(), Date.now(), requestAnimationFrame, setTimeout—all seeked frame by frame.

Same composition = same video. Every time.

3. Verify

Six gates machine-check the result before delivery:

A failing video comes back marked FAILED, with the gate report and an instruction not to deliver it, so your AI fixes the composition and renders again.

4. See

envie_see returns frames at chosen timestamps so your AI can judge layout and pacing—not just whether the file exists.

5. Translate

envie_translate reads the finished file as data: per-frame motion, every cut and fade, still holds, LUFS, true peak, and how each audio hit sits against the nearest picture event.

Requirements

Audio options

Runs entirely on your machine. Nothing reaches GOL servers.

Formats

Shortened. The full README is on GitHub.